High-resolution computed tomography assessment of connective tissue disease-associated interstitial lung disease: Advances in diagnosis and prognostic evaluation

Abstract

Connective tissue disease-associated interstitial lung disease (CTD-ILD), an important pulmonary manifestation of systemic autoimmune diseases and a major determinant of morbidity and mortality, is highly heterogeneous across disease subtypes, requiring reliable imaging assessment. High-resolution computed tomography (HRCT) remains the cornerstone imaging modality in this setting. This narrative review summarizes recent advances in HRCT-based assessment of CTD-ILD, based on a structured literature search of PubMed/MEDLINE, EMBASE, and Web of Science through May 31, 2026, with emphasis on diagnosis, prognostic assessment, and longitudinal monitoring. Conventional visual interpretation and semiquantitative scoring remain the established foundation of clinical assessment. Quantitative computed tomography provides an emerging objective adjunct but remains variably implemented across centers, whereas radiomics, machine learning, and deep learning offer promising capabilities for automated segmentation, high-dimensional phenotyping, pattern classification, and individualized risk prediction but remain predominantly investigational. Current evidence supports a layered imaging framework in which conventional and advanced approaches are complementary rather than competing; however, the maturity and strength of evidence vary substantially across CTD subtypes and clinical applications. Relatively more evidence is available for systemic sclerosis-associated ILD, rheumatoid arthritis-associated ILD, and inflammatory myopathy-associated ILD, whereas evidence for Sjögren syndrome-associated ILD and systemic lupus erythematosus-associated ILD remains comparatively limited. Future research should prioritize standardized imaging workflows, reproducibility, harmonized endpoints, prospective multicenter validation, and demonstration of clinical utility before broader implementation of advanced HRCT-derived biomarkers.

Key Words: Connective tissue disease-associated interstitial lung disease; High-resolution computed tomography; Semiquantitative scoring; Quantitative computed tomography; Radiomics; Machine learning

Core Tip: This narrative review summarizes current advances in high-resolution computed tomography (CT)-based assessment of connective tissue disease-associated interstitial lung disease across diagnosis, prognostic assessment, and longitudinal monitoring. Visual interpretation and semiquantitative scoring remain the established foundation of clinical practice, whereas quantitative CT provides an emerging objective adjunct. Radiomics, machine learning, and deep learning offer promising capabilities but remain predominantly investigational. We propose a layered, complementary framework that reflects current clinical readiness rather than a validated clinical pathway and highlight key priorities for future translation, including standardization, reproducibility, multicenter validation, harmonized endpoints, and demonstration of clinical utility.

INTRODUCTION

Connective tissue disease-associated interstitial lung disease (CTD-ILD) represents one of the most clinically important pulmonary manifestations of systemic autoimmune rheumatic diseases[1]. It may occur in systemic sclerosis (SSc), rheumatoid arthritis (RA), Sjögren syndrome, idiopathic inflammatory myopathy (IIM), systemic lupus erythematosus (SLE), and other connective tissue diseases; however, its prevalence, radiologic pattern, clinical course, and treatment response substantially differ across disease subtypes[1-4]. In some patients, lung involvement remains mild and stable for years, whereas in others, it rapidly progresses to extensive fibrosis, respiratory failure, and death. This heterogeneity makes early recognition, accurate phenotyping, and reliable prognostic assessment essential for clinical management[5-7].

High-resolution computed tomography (HRCT) remains the cornerstone imaging modality for CTD-ILD detection, characterization, and follow-up, with recent international recommendations supporting the central role of HRCT together with pulmonary function tests (PFTs) in screening and longitudinal assessment of CTD-ILD[8-10]. Visual interpretation and semiquantitative scoring provide accessible and clinically interpretable assessments of disease pattern, extent, and severity, which may support diagnosis and risk stratification and, when integrated with clinical and pulmonary function data, inform treatment-related decision-making[11]. However, reader experience and interobserver variability inherently influence visual assessment, which may be less sensitive to subtle abnormalities or small longitudinal changes. These limitations are particularly relevant in early disease and in patients with mixed inflammatory and fibrotic manifestations, where accurate characterization and monitoring remain challenging[12,13]. To address these challenges, HRCT assessment has evolved beyond conventional visual interpretation toward more quantitative and computational approaches. Quantitative computed tomography (CT) enables the objective measurement of lung attenuation, volume, texture, and disease extent, whereas radiomics, machine learning, and deep learning techniques facilitate high-dimensional feature extraction, automated pattern recognition, and predictive modeling[14-17]. Rather than replacing visual interpretation, these approaches complement traditional assessment within a layered HRCT-based framework. Previous reviews have primarily addressed conventional HRCT manifestations of CTD-ILD, quantitative CT methodology, or individual computational approaches such as radiomics and artificial intelligence, including studies extending across broader ILD populations[1,13-17]. The present review therefore adopts a different, clinically oriented perspective by integrating conventional and computational HRCT approaches according to three major clinical tasks: Diagnosis, prognostic assessment, and longitudinal monitoring. Particular emphasis is placed on the maturity and clinical readiness of each approach, distinguishing established clinical methods from emerging adjunctive techniques and predominantly investigational computational tools. The proposed layered framework should therefore be understood as a conceptual synthesis of the current evidence rather than a prospectively validated clinical pathway. The primary objective of this narrative review is to synthesize current HRCT-based approaches for CTD-ILD across diagnosis, prognostic assessment, and longitudinal monitoring, with particular emphasis on their evidence maturity, clinical readiness, and priorities for future translation. By organizing conventional and computational imaging methods within a unified clinical-task framework, this review aims to bridge the gap between technical development and practical interpretation of their current and potential roles in CTD-ILD assessment.

LITERATURE SEARCH STRATEGY

This narrative review was based on a literature search of PubMed/MEDLINE, EMBASE, and Web of Science from database inception to May 31, 2026. The search was intended to support a narrative synthesis rather than a systematic review or meta-analysis. Accordingly, no original statistical analyses, hypothesis testing, or quantitative pooling of study results were performed. Boolean combinations of free-text terms related to “connective tissue disease-associated interstitial lung disease”, “CTD-ILD”, “high-resolution computed tomography”, “HRCT”, “quantitative CT”, “semiquantitative scoring”, “radiomics”, “machine learning”, “deep learning”, and individual connective tissue diseases, including systemic sclerosis, rheumatoid arthritis, idiopathic inflammatory myopathy, Sjögren syndrome, and systemic lupus erythematosus, were used.

For the purposes of this narrative review, eligible publications included English-language original studies, relevant reviews, international guidelines, and consensus statements addressing HRCT-based diagnosis, prognostic assessment, longitudinal monitoring, or treatment-response evaluation in CTD-ILD. Non-English-language publications and publications outside the scope of HRCT-based assessment of CTD-ILD were excluded. Priority was given to CTD-ILD-specific studies, multicenter cohorts, externally validated models, and recent studies, while seminal earlier studies establishing clinically relevant HRCT concepts or scoring systems were also retained. Studies of non-CTD interstitial lung disease were considered only when they provided directly relevant methodological evidence for HRCT-based quantitative or computational assessment. Reference lists of selected publications were additionally screened to identify relevant studies that might not have been captured by the primary database searches.

Study selection was performed jointly by Jiang SY and Su KX through discussion of the relevance of each publication to the scope of the review and its contribution to the evidence synthesis. When consensus could not be reached, Li R participated in the discussion and made the final decision regarding inclusion. Preprints were considered only when they were directly relevant to the review topic and were explicitly identified in the manuscript as non-peer-reviewed evidence. For studies reporting diagnostic, prognostic, or longitudinal performance, key quantitative information, including sample size, follow-up duration, area under the receiver operating characteristic curve, concordance index (C-index), thresholds, and principal outcome measures, was manually checked against the original full-text publication or published abstract. Because this review was designed as a narrative synthesis rather than a systematic review, no formal PRISMA-based study-selection process or risk-of-bias assessment tool was applied. Accordingly, despite the structured search and study-prioritization process, selective searching and restriction to English-language publications may have resulted in the omission of some relevant evidence.

BIOLOGICAL BASIS OF IMAGING HETEROGENEITY IN CTD-ILD

The marked heterogeneity of CTD-ILD on HRCT reflects the diversity of its underlying biological mechanisms[18]. Pathogenic pathways differ across connective tissue disease subtypes; however, CTD-ILD is generally considered the pulmonary consequence of interacting immune-inflammatory, vascular, epithelial, and profibrotic processes. Genetic susceptibility and environmental exposures may trigger repeated alveolar epithelial or endothelial injury in the setting of systemic autoimmunity[19,20]. This injury promotes innate and adaptive immune activation, the release of inflammatory and profibrotic mediators, abnormal epithelial-mesenchymal and endothelial-mesenchymal crosstalk, fibroblast activation, myofibroblast differentiation, and excessive extracellular matrix deposition[21,22]. Compared with idiopathic pulmonary fibrosis (IPF), dysregulated inflammatory and autoimmune pathways play a more prominent initiating role in CTD-ILD, despite several overlapping downstream fibrotic programs[23,24]. Hence, CTD-ILD may present with variable combinations of inflammatory infiltration, organizing pneumonia (OP), ground-glass opacity (GGO), reticulation, traction bronchiectasis, and established fibrosis[25].

Subtype-specific mechanisms further contribute to differences in imaging phenotype and clinical behavior. Endothelial injury, microvascular dysfunction, defective vascular repair, and fibroblast activation are central to progressive fibrotic remodeling in SSc-associated ILD, which commonly manifests as lower-lung-predominant GGO, reticulation, and traction bronchiectasis[26,27]. Smoking-related airway and mucosal injury, aberrant protein citrullination, autoantibody production, and genetic susceptibility, including MUC5B promoter variants, are closely associated with fibrotic lung involvement and a usual interstitial pneumonia (UIP)-like phenotype in some patients with RA-associated ILD[28,29]. Myositis-specific autoantibodies are strongly associated with lung involvement in IIM-associated ILD, particularly anti-synthetase syndrome and anti-melanoma differentiation-associated gene 5 (anti-MDA5)-positive disease, and anti-MDA5 positivity is frequently related to rapidly progressive ILD characterized by extensive GGO and consolidation[30-33].

Therefore, CTD-ILD should not be viewed as a single uniform fibrotic disorder. Rather, it represents a spectrum of immune-mediated lung injuries with different degrees of inflammation, vascular damage, epithelial injury, and fibrosis[18,34]. This biological diversity provides the rationale for HRCT-based assessment beyond the simple detection of ILD. Visual pattern recognition, semiquantitative extent scoring, quantitative CT metrics, radiomics, and deep learning approaches may capture different aspects of heterogeneity, including inflammatory burden, fibrotic remodeling, spatial distribution, and longitudinal structural change[12,35-37]. Therefore, understanding the biological basis of imaging heterogeneity is essential for interpreting HRCT findings and integrating imaging features with clinical, serological, and pulmonary function data in diagnosis, prognostic assessment, and longitudinal monitoring[38].

FRAMEWORK OF HRCT-BASED QUANTITATIVE AND COMPUTATIONAL ASSESSMENT IN CTD-ILD

HRCT-based assessment in CTD-ILD has evolved from descriptive visual interpretation toward a layered framework that integrates conventional, quantitative, and computational approaches. This evolution does not imply replacing expert radiologic assessment. Rather, each method addresses a different level of clinical and imaging information: Visual interpretation defines the dominant morphologic pattern, semiquantitative scoring estimates disease extent and severity, quantitative CT provides standardized numerical descriptors, radiomics and machine learning capture high-dimensional imaging heterogeneity, and deep learning supports automation and serial image analysis[12,14,17,36,39].

Visual interpretation remains the foundation of routine HRCT assessment because it enables direct recognition of key parenchymal abnormalities, including GGO, consolidation, reticulation, traction bronchiectasis, honeycombing, and OP-like changes[40]. Further, semiquantitative scoring systems translate these findings into structured estimates of disease burden, providing clinically interpretable information for baseline assessment, risk stratification, and follow-up comparison[11,39]. Their main strength lies in accessibility and interpretability; however, reader experience, interobserver variability, and the difficulty of visually detecting subtle or diffuse interval changes inevitably influence their performance.

Quantitative CT complements visual assessment by converting lung abnormalities into reproducible numerical measures. These measures may include lung volume, attenuation distribution, density histogram parameters, texture burden, and the proportion of specific parenchymal patterns, including GGO, reticulation, honeycombing, or quantitative lung fibrosis, depending on the software and analytic strategy. These metrics are particularly useful in diffuse disease, early lung involvement, and longitudinal monitoring, where small structural changes may be challenging to judge visually[41]. However, quantitative CT is sensitive to acquisition parameters, reconstruction algorithms, inspiratory effort, segmentation quality, and postprocessing workflows, which limit direct comparability across institutions[42-44].

Radiomics, machine learning, and deep learning further extend HRCT assessment from predefined measurements to data-driven characterization of imaging heterogeneity. Radiomics extracts high-dimensional features associated with intensity, shape, spatial distribution, and texture, which can be incorporated into machine learning models for classification, phenotyping, progression prediction, and integrated clinical-imaging risk stratification[36,45]. Deep learning methods can be employed to additionally learn hierarchical imaging representations directly from CT data and have particular value for automated lung and lesion segmentation, pattern recognition, disease-extent quantification, and longitudinal image comparison[12,46]. However, the clinical translation of these approaches remains constrained by small datasets, limited external validation, variable CTD-ILD populations, and insufficient model interpretability[14,47].

From the perspective of clinical readiness, these approaches should not be regarded as equally mature. Conventional visual interpretation remains the established foundation of routine CTD-ILD assessment, while semiquantitative scoring provides a clinically interpretable adjunct for structured assessment of disease extent, severity, and longitudinal change. Quantitative CT offers more objective measurements and has accumulated increasing validation, but its routine implementation remains variable because of differences in acquisition protocols, software platforms, postprocessing workflows, and local expertise. In contrast, radiomics, machine learning, and deep learning remain predominantly investigational in CTD-ILD. Although these approaches have demonstrated promising performance for segmentation, classification, and risk prediction, much of the available evidence is derived from retrospective or selected cohorts, with limited prospective multicenter validation and clinical-utility assessment. Accordingly, this hierarchy reflects the current maturity of evidence and should not be interpreted as a ready-to-use clinical workflow.

Altogether, HRCT-based assessment of CTD-ILD should be understood as a complementary rather than competitive framework. Visual and semiquantitative methods provide clinical interpretability, quantitative CT improves objectivity and reproducibility, radiomics and machine learning refine phenotyping and prediction, and deep learning facilitates automation and longitudinal analysis. This layered framework provides the methodological basis for diagnostic, prognostic, and monitoring applications discussed in the following sections. A detailed comparison of these HRCT-based assessment approaches, including their core characteristics, main advantages, clinical applications, validation status, and current limitations, is summarized in Table 1. Representative study-level evidence across diagnostic, prognostic, and longitudinal applications, including study population, HRCT approaches, principal findings, and key limitations, is summarized in Table 2.

Table 1 Comparison of high-resolution computed tomography-based assessment approaches in connective tissue disease-associated interstitial lung disease.

Approach	Core characteristics	Main advantages	Main clinical applications	Current evidence	Main limitations
Visual interpretation	Direct radiologist assessment of recognizable HRCT patterns and abnormalities, including GGO, consolidation, reticulation, traction bronchiectasis, honeycombing, and OP-like changes; also allows integrated assessment of lesion distribution and dominant morphologic pattern	Widely available; rapid; does not require dedicated software; directly linked to established radiologic terminology and multidisciplinary clinical decision-making; preserves morphologic context that may be difficult to fully capture with numerical features	Initial detection and characterization of CTD-ILD; identification of dominant HRCT pattern; assessment of inflammatory vs fibrotic manifestations; qualitative evaluation of disease progression and treatment-related change; support for multidisciplinary diagnosis	Most established approach in routine practice and supported by extensive CTD-ILD literature across multiple disease subtypes. It remains the clinical reference framework against which many quantitative and computational methods are compared[39,40]	Dependent on reader experience and pattern definitions; subject to interobserver variability; subtle, diffuse, or mixed abnormalities may be difficult to assess consistently; small longitudinal changes can be challenging to recognize and quantify reproducibly
Semiquantitative scoring	Structured visual grading of disease extent or severity using predefined lobar, zonal, or whole-lung scoring systems for GGO, fibrosis, reticulation, honeycombing, or total ILD burden	More structured and reproducible than purely descriptive visual assessment; provides clinically interpretable estimates of disease burden; relatively easy to apply without complex computational infrastructure; facilitates baseline risk stratification and serial comparison	Quantification of baseline disease extent; severity assessment; risk stratification; evaluation of inflammatory and fibrotic burden; prognostic assessment; longitudinal monitoring and treatment-response evaluation	Well established in CTD-ILD, particularly in SSc-ILD and inflammatory myopathy-associated ILD, and frequently used in observational cohorts and clinical studies. Several scoring systems have demonstrated associations with pulmonary function and clinical outcomes[11,39,69]	Still relies on reader judgment; scoring definitions and anatomic sampling schemes differ among studies; ceiling/floor effects and coarse categorical scales may reduce sensitivity to subtle changes; reproducibility can decline in heterogeneous or mixed-pattern disease
Quantitative CT	Computer-assisted extraction of predefined numerical measures from HRCT, such as lung volume, attenuation distribution, density histogram parameters, texture burden, and the proportion or extent of specific parenchymal abnormalities	Provides continuous and more objective measurements of disease burden; may improve sensitivity to diffuse or subtle abnormalities; enables more standardized longitudinal comparison than visual estimation alone; can quantify structural change that may be difficult to appreciate visually	Objective measurement of ILD extent and severity; quantification of fibrosis, GGO, reticulation, and other parenchymal patterns; early disease detection; correlation with pulmonary function; prognostic assessment; longitudinal progression and treatment-response monitoring	Quantitative CT has accumulated an increasing body of evidence in SSc-ILD and has also been investigated in RA-ILD, inflammatory myopathy-associated ILD, and mixed CTD-ILD cohorts. It represents an emerging objective adjunct, although implementation remains variable across centers. Some quantitative metrics have also been evaluated in longitudinal cohorts and treatment studies[41,79,91,100]	Sensitive to scanner type, acquisition parameters, reconstruction algorithms, inspiratory effort, segmentation quality, and postprocessing software; results may not be directly comparable across institutions; platform-specific outputs and limited standardization can restrict reproducibility
Radiomics	High-dimensional extraction of quantitative features describing image intensity, shape, spatial distribution, and texture, usually followed by feature selection and construction of a radiomics score or integrated clinical-imaging model	Captures imaging heterogeneity beyond conventional visual and predefined quantitative measures; can identify subtle patterns not readily visible to the human eye; supports combination of imaging features with clinical, serologic, or functional variables for individualized prediction	Detection of CTD-ILD or subclinical lung involvement; disease phenotyping; classification of imaging patterns; assessment of disease severity; risk stratification; prediction of progression, rapidly progressive ILD, mortality, or other clinical outcomes	Promising CTD-ILD-specific evidence is available, including studies with internal and external validation in inflammatory myopathy-associated ILD and RA-ILD. However, the overall evidence base remains smaller and less standardized than that for conventional and semiquantitative assessment, and radiomics remains predominantly investigational[36,45,71,83]	Feature values are sensitive to acquisition, reconstruction, segmentation, and preprocessing choices; high-dimensional features increase the risk of overfitting in small datasets; feature reproducibility and interpretability may be limited; many studies remain retrospective and external validation is still inconsistent
Machine learning	Use of statistical or computational algorithms to learn relationships among imaging, clinical, physiologic, and serologic variables for classification or outcome prediction; commonly includes methods such as logistic regression-based learning, random forest, support vector machine, and gradient-boosting approaches	Can integrate multiple heterogeneous variables and model complex nonlinear relationships; allows development of combined clinical-imaging prediction tools; may improve individualized diagnostic or prognostic discrimination compared with single-variable approaches	Diagnostic classification; differentiation of CTD-ILD phenotypes; integrated risk prediction; prediction of progression, pulmonary function decline, rapidly progressive disease, mortality, or other clinically relevant outcomes	Several CTD-ILD studies have reported favorable diagnostic or prognostic performance, including multicenter analyses in selected settings. Nevertheless, many models are developed retrospectively, independent multicenter external validation remains limited across the broader CTD-ILD spectrum, and machine-learning approaches remain predominantly investigational[62,67,85]	Performance strongly depends on cohort size, variable quality, and model-development strategy; susceptible to overfitting and optimistic performance estimates in small or selected datasets; generalizability across institutions and CTD subtypes may be limited; implementation requires reproducible input variables and standardized workflows
Deep learning	Neural-network-based learning of hierarchical image representations directly from CT data, reducing reliance on predefined handcrafted features; commonly applied to automated segmentation, pattern recognition, quantitative mapping, and serial image analysis	Supports automated or semi-automated image analysis; can reduce manual segmentation burden; enables direct recognition of complex spatial patterns; particularly suitable for whole-lung or lesion segmentation, automated pattern classification, and longitudinal comparison	Automated lung and lesion segmentation; quantification of ILD extent and individual HRCT patterns; pattern classification; correlation of imaging burden with physiologic impairment; prognostic assessment; detection and quantification of longitudinal structural change	Feasibility and good performance have been demonstrated in SSc-ILD, inflammatory myopathy-associated ILD, and mixed CTD-ILD cohorts, including studies using independent test or validation datasets. However, validation remains variable, and deep-learning approaches remain predominantly investigational rather than established tools for routine clinical use[12,46,48,87]	Requires sufficiently large and well-annotated datasets and greater computational resources; performance may vary across scanners, protocols, and patient populations; training and maintenance can be resource intensive; external validation and standardized deployment remain insufficient for widespread routine use

HRCT: High-resolution computed tomography; GGO: Ground-glass opacity; OP: Organizing pneumonia; CTD-ILD: Connective tissue disease-associated interstitial lung disease; SSc-ILD: Systemic sclerosis-associated interstitial lung disease; RA-ILD: Rheumatoid arthritis-associated interstitial lung disease.

Table 2 Representative high-resolution computed tomography-based studies in connective tissue disease-associated interstitial lung disease by clinical application.

Clinical application	Ref.	CTD subtype/population	HRCT approach	Main finding	Main limitation
Diagnosis	Frauenfelder et al[50]	SSc	Reduced 9-slice visual HRCT	Sensitivity: 88.3%; radiation dose decreased from 2.09 ± 1.34 mSv to 0.08 ± 0.06 mSv	SSc-specific reduced-slice screening may miss limited abnormalities
	Wang et al[51]	Juvenile DM-ILD	Semiquantitative CT scoring	CT score correlated strongly with serum KL-6 (r = 0.784, P < 0.01)	Pediatric disease-specific population limits generalizability
	Chen et al[49]	CTD-ILD	Dual-energy quantitative CT	Whole-lung monochromatic CT number identified extensive disease (AUC: 0.901; sensitivity: 82.1%; specificity: 85.4%)	Cross-sectional design; prognostic relevance not established
	Yang et al[62]	PM/DM	Quantitative CT + machine learning	Random forest achieved AUC 0.843 in the temporal test cohort for ILD detection.	Retrospective study; independent external validation is needed
	Xu et al[63]	IIM	Clinical + CT radiomics machine learning	Combined nomogram achieved AUCs of 0.877 and 0.898 in training and testing cohorts	Independent external validation and prospective assessment are needed
Prognostic assessment	Goh et al[69]	SSc-ILD	Visual HRCT extent + FVC	HRCT extent > 20% defined extensive disease; FVC 70% aided classification when HRCT extent was 10%-30%	Developed for SSc-ILD; generalizability to other CTD-ILD is uncertain
	Schniering et al[71]	SSc-ILD	Quantitative radiomic risk score	High-risk status predicted shorter progression-free survival and increased progression risk (HR: 5.14) after external validation	Radiomics workflow requires standardization and broader prospective validation
	Xu et al[83]	Anti-MDA5-positive DM-ILD	Radiomics + age + FVC	Rad-score plus model achieved C-indices of 0.88, 0.88, 0.83, and 0.84 across four datasets	Disease-specific model; generalizability beyond anti-MDA5-positive DM-ILD is uncertain
Longitudinal monitoring and treatment response	Goldin et al[91]	SSc-ILD	Volumetric QILD	Whole-lung QILD decreased after treatment and its change correlated with FVC, DLCO, and dyspnea score	Imaging change alone should not be interpreted as proof of treatment efficacy
	Occhipinti et al[100]	SSc-ILD	Quantitative CT during immunosuppressive therapy	Total lung-volume change predicted a composite FVC/DLCO endpoint (AUC: 0.74); individual texture changes had limited predictive ability	Small cohort; individual texture patterns had limited predictive ability
	Tian et al[114]	IIM-ILD	Quantitative fibrosis growth-rate modeling	Five longitudinal patterns were identified; rapid progression was significantly associated with mortality	Small exploratory cohort; the growth-rate model requires further validation

CTD: Connective tissue disease; HRCT: High-resolution computed tomography; SSc: Systemic sclerosis; DM-ILD: Dermatomyositis-associated interstitial lung disease; KL-6: Krebs von den Lungen-6; AUC: Area under the receiver operating characteristic curve; PM: Polymyositis; IIM: Idiopathic inflammatory myopathy; CT: Computed tomography; FVC: Forced vital capacity; HR: Hazard ratio; Anti-MDA5: Anti-melanoma differentiation-associated gene 5; QILD: Quantitative interstitial lung disease; DLCO: Diffusing capacity of the lung for carbon monoxide.

HRCT-BASED DIAGNOSTIC ASSESSMENT IN CTD-ILD

HRCT remains the central imaging modality for detecting lung involvement, characterizing dominant radiologic patterns, and estimating baseline disease extent in the diagnostic evaluation of CTD-ILD. However, diagnostic assessment should not be simply understood as confirming the presence or absence of ILD. In clinical practice, HRCT-based diagnosis also involves inflammatory and fibrotic manifestation recognition, disease burden evaluation, subtle or subclinical abnormality identification, and interpretation standardization across readers and institutions. Within this context, visual interpretation and semiquantitative scoring remain the practical foundation, whereas quantitative CT, radiomics, machine learning, and deep learning provide complementary tools for objective, reproducible, and data-driven evaluation[48-50].

Visual interpretation and semiquantitative scoring

Visual interpretation and semiquantitative scoring remain the most accessible diagnostic approaches in CTD-ILD because they directly reflect clinically recognizable HRCT abnormalities, including GGO, consolidation, reticulation, traction bronchiectasis, honeycombing, and OP-like changes. Their diagnostic value is particularly evident in high-risk CTD populations, where structured HRCT-based assessment can support early detection while maintaining clinical interpretability.

In SSc, Frauenfelder et al[50] prospectively examined a reduced 9-slice HRCT protocol in 170 analyzable patients. Compared with full whole-chest HRCT, the reduced protocol detected ILD in 68 of 77 cases, demonstrating a sensitivity of 88.3% for both readers, while substantially reducing radiation exposure from 2.09 ± 1.34 mSv to 0.08 ± 0.06 mSv. This study revealed that structured visual HRCT assessment can be adapted for screening in high-risk CTD populations, particularly when repeated imaging or radiation exposure is a concern.

Semiquantitative scoring provides a more structured way to describe the extent and severity of visually identified CT abnormalities. Wang et al[51] developed a semiquantitative CT scoring system for juvenile dermatomyositis (DM)-associated ILD and reported a strong correlation between CT score and serum Krebs von den Lungen-6 (KL-6) level (r = 0.784, P < 0.01). Kobayashi et al[52] similarly found increased serum KL-6 levels in juvenile DM patients with ILD, which correlated with the degree of CT abnormalities. These studies were conducted in specific pediatric populations; however, they support the broader concept that structured visual scoring can reflect clinically relevant pulmonary involvement and may serve as an interpretable baseline assessment tool.

However, visual and semiquantitative approaches have inherent limitations. Their performance relies on reader experience, scoring definitions, and the ability to visually estimate diffuse or subtle abnormalities. These limitations become particularly important in early CTD-ILD, mixed inflammatory and fibrotic disease, and follow-up assessment, where small changes in disease extent may be clinically meaningful but challenging to judge consistently.

Quantitative CT

Quantitative CT complements visual interpretation by converting parenchymal abnormalities into numerical measures of attenuation, lung volume, texture, vascular parameters, airway-related changes, and disease burden. Its value in diagnostic assessment extends across several levels: Detecting early or clinically silent pulmonary involvement, supporting objective phenotyping and pattern classification, and quantifying disease extent and severity in association with functional impairment.

Early or subclinical pulmonary involvement: Quantitative CT may help identify early or clinically silent pulmonary involvement beyond overt ILD detection. Hoffmann et al[53] reported that approximately one-quarter of treatment-naïve patients with newly diagnosed CTD-ILD were asymptomatic at initial CTD diagnosis; however, even asymptomatic patients showed reduced diffusing capacity of the lung for carbon monoxide (DLCO) and quantifiable interstitial abnormalities on artificial intelligence (AI)-based quantitative HRCT. This finding highlights the potential mismatch between symptoms and objective pulmonary involvement in early CTD-ILD.

Quantitative CT may also detect abnormalities before overt ILD becomes visible on conventional HRCT. Hei et al[54] found significantly reduced standardized pulmonary vascular volume in patients with polymyositis or DM (PM/DM) without visually apparent ILD compared with healthy controls. The standardized vascular volume of the right middle lobe achieved an area under the curve (AUC) of 0.870 for differentiating patients with PM/DM from controls. These data indicate that quantitative CT may capture subtle pulmonary alterations before they are recognized visually. However, the clinical significance and prognostic implications of such subclinical abnormalities require further validation.

Diagnostic phenotyping and pattern classification: After detecting early lung involvement, quantitative CT may further support diagnostic phenotyping across CTD subtypes and imaging patterns. In IIM-ILD, Zhang et al[55] used dual-layer detector spectral CT to compare patients with antisynthetase syndrome and PM/DM. Spectral parameters, including lung volume, monochromatic CT number, effective atomic number, and electron density, differed between groups, particularly in the lower lobes. These results suggest that quantitative spectral CT may provide objective information for distinguishing patterns of lung involvement across IIM subtypes.

Quantitative CT also integrates density, airway, and vascular information for CTD-ILD identification. Cheng et al[56] demonstrated that a quantitative model combining lung density, small airway, and pulmonary vascular parameters achieved high diagnostic performance for CTD-associated ILD, with an AUC of 0.963. Although this approach requires further validation, it indicates that quantitative CT characterizes not only parenchymal abnormalities but also airway and vascular alterations associated with CTD-associated lung involvement.

The evidence remains partly indirect for radiologic pattern classification. Chikasue et al[57] applied whole-lung volume histogram analysis in interstitial pneumonia cohorts that included CTD-associated nonspecific interstitial pneumonia (NSIP) and reported AUCs of 0.91 in the derivation cohort and 0.81 in the external validation cohort to identify a UIP pattern. The distinction between UIP-like fibrotic disease, NSIP, and inflammatory-predominant patterns has diagnostic and prognostic implications in CTD-ILD; thus, histogram-based analysis may provide objective support for pattern classification. However, such evidence should be interpreted as methodologically supportive rather than definitive CTD-ILD-specific diagnostic validation.

Disease extent, severity, and functional correlation: After establishing lung involvement and radiologic phenotype, quantitative CT can provide a more standardized assessment of disease extent and severity. In a cross-sectional study of 147 patients with CTD-ILD, Chen et al[49] used dual-energy CT to assess the effective atomic number, monochromatic CT number, and lung volume. Whole-lung monochromatic CT number demonstrated the best discrimination for extensive CTD-ILD, with an AUC of 0.901, sensitivity of 82.1%, and specificity of 85.4% at a cutoff of -762.30 Hounsfield units (HU). These findings indicate that quantitative attenuation-based parameters may help standardize the assessment of extensive disease, particularly when diffuse or heterogeneous involvement limits visual estimation.

Quantitative CT also shows clinically meaningful correlations with visual scoring and pulmonary function. In SSc-ILD, Muangchan et al[58] developed and internally validated a semiautomated HRCT-based quantification tool in 78 patients. The programme demonstrated excellent intra- and interobserver reproducibility, with intraclass correlation coefficients exceeding 0.9 for lung segmentation and lesion quantification. Programme-derived global ILD extent correlated strongly with expert-assessed extent (r = 0.74, P < 0.0001), and a programme-derived threshold of 19.5% identified expert-assessed ILD extent > 20% with an AUC of 0.885. Similarly, AI-based quantitative HRCT studies by Hoffmann et al[59] demonstrated that quantitative measures of GGO, reticulation, high-attenuation areas, and total ILD volume were correlated with pulmonary function parameters, including DLCO and total lung capacity. Collectively, these results support quantitative CT as an objective measure of disease burden with clinically relevant associations with visual scores and pulmonary function.

Emerging CT technologies: Emerging CT technologies may further improve quantitative diagnostic assessment. In a propensity score-matched SSc cohort, Happe et al[60] compared automated lung texture analysis between photon-counting detector CT and conventional energy-integrating detector CT. Photon-counting detector CT achieved higher AUCs for detecting ILD and specific features, including reticulation and honeycombing, while substantially reducing radiation dose. Although still preliminary, advances in CT hardware may improve the feasibility of repeated low-dose quantitative HRCT assessment in CTD-ILD.

Radiomics and machine learning

Radiomics and machine learning extend diagnostic HRCT assessment by extracting high-dimensional imaging features that are not readily captured by visual inspection or conventional quantitative metrics. Their diagnostic value in CTD-ILD can be considered in two major scenarios: Distinguishing inflammatory from fibrotic disease and constructing CTD-specific diagnostic models for ILD detection or early disease recognition.

Inflammatory-fibrotic pattern discrimination: Pattern discrimination is clinically relevant because CTD-ILD frequently contains overlapping inflammatory and fibrotic components. Colligiani et al[61] analyzed patients with ILD, including cellular NSIP secondary to connective tissue disease and IPF, and incorporated 103 radiomic features into an explainable machine-learning framework. The model achieved high diagnostic performance for distinguishing fibrosing from inflammatory ILD patterns, with an area under the receiver operating characteristic curve of 0.98. Although the study population was not restricted to CTD-ILD, the results provide methodological support for using radiomics to distinguish inflammatory-predominant from fibrotic-predominant patterns, a distinction that may be relevant to CTD-ILD phenotyping and potentially to treatment-related assessment.

CTD-specific diagnostic models and early disease detection: More direct evidence comes from CTD-specific diagnostic models. Yang et al[62] conducted a retrospective study of 348 patients with PM/DM and extracted quantitative CT parameters, including lung volume, mean lung density, intrapulmonary vascular volume, and high-attenuation area percentage. Among the seven machine-learning classifiers, the random forest model achieved the best performance in the temporal test cohort, with an AUC of 0.843. Xu et al[63] developed an interpretable machine learning model integrating clinical and CT radiomic features for detecting ILD in IIM. The combined nomogram achieved AUCs of 0.877 and 0.898 in the training and testing cohorts, respectively, outperforming clinical-only and radiomics-only models. In RA-ILD, Wu et al[64] developed machine-learning models integrating multi-omics and radiomics data in 278 patients with RA from two cohorts. The combined imaging-clinical logistic regression model achieved AUCs of 0.963 and 0.913 in the internal and external validation cohorts, respectively. These studies suggest that integrated models combining HRCT-derived features with clinical or multi-omics information may improve diagnostic classification. However, the high performance reported in relatively small or retrospective datasets should be interpreted cautiously until external validation.

Radiomics may also facilitate early disease detection. Liu et al[65] evaluated children with juvenile DM using automated lung delineation and least absolute shrinkage and selection operator (LASSO)-selected radiomic features. A radiomics score combined with anti-MDA5 antibody status achieved AUCs of 0.92 and 0.93 in the training and validation cohorts, respectively, for identifying juvenile DM-associated ILD. Notably, some patients not initially recognized as having ILD on conventional HRCT interpretation already demonstrated elevated baseline radiomics scores. This observation supports the potential value of radiomics in detecting mild or diffuse abnormalities that may escape routine visual assessment, despite requiring prospective confirmation to identify whether these early signals predict clinically meaningful disease.

Deep learning and AI-assisted standardization

Deep learning and AI-assisted texture analysis provide additional diagnostic value by improving automation, segmentation consistency, pattern classification, and comparison with expert visual assessment. These methods are particularly attractive in CTD-ILD because disease abnormalities may be spatially heterogeneous, subtle, and challenging to reproducibly quantify across readers.

Automated segmentation, quantification, and functional relevance: Chassagnon et al[48] developed a deep learning algorithm for SSc-ILD using CT scans from 208 patients for the automated delineation of ILD extent. Agreement between the algorithm and radiologists, measured using the Dice similarity coefficient, ranged from 0.74 to 0.75, compared with 0.68-0.71 between radiologists, indicating that automated contouring can achieve performance comparable to expert assessment. Further, Esnaashari et al[66] applied a U-Net-based model to segment GGO and reticulation in SSc, achieving Dice coefficients of 87.22% for reticulation and 86.20% for combined GGO and reticulation. These studies support the feasibility of automated segmentation and quantification, despite the relatively small number of annotated patients in some datasets, which remains an important limitation.

Deep learning quantification may also improve the functional relevance of the baseline diagnostic assessment. Ito et al[12] assessed a deep learning quantification tool in 80 patients with CTD-ILD and revealed that deep learning-derived quantitative interstitial lung disease (QILD) scores identified patients with forced vital capacity (FVC) of < 70% predicted more accurately than visual scoring, with AUCs of 0.833 vs 0.660. The higher diagnostic performance of automated quantification suggests that it may capture physiologically relevant disease burden more effectively than conventional visual estimation alone.

Pattern classification and agreement with expert assessment: Zhang et al[67] developed a deep learning model for HRCT pattern classification in 629 patients with IIM-ILD to classify five imaging patterns: NSIP, OP, NSIP + OP overlap, UIP, and diffuse alveolar damage. The model achieved a mean AUC of 0.885 and 0.835 in the internal test and external validation cohorts, respectively, supporting its potential as a diagnostic support tool for complex myositis-associated ILD patterns. Such models may be particularly useful when multiple patterns coexist or when radiologic categorization is difficult in routine practice.

AI-based texture analysis enables comparison of automated quantification with expert visual assessment. Kuzmanovic et al[68] evaluated 107 patients with IIM-ILD and revealed that AI-derived scores correlated most strongly with radiologist assessment for normal lung, GGO, and consolidation, whereas agreement was weaker for reticulation and honeycombing. This pattern is clinically important: AI tools may be more reliable for relatively homogeneous inflammatory abnormalities than for complex fibrotic patterns. Therefore, AI-assisted analysis should currently be considered a complementary tool to expert interpretation rather than a standalone diagnostic replacement.

Summary and current evidence gaps

Current evidence supports a layered diagnostic framework for HRCT-based assessment in CTD-ILD. Visual interpretation and semiquantitative scoring remain indispensable because they are accessible, clinically interpretable, and directly associated with routine radiologic decision-making. Quantitative CT improves objectivity in detecting early or diffuse involvement, supporting diagnostic phenotyping, and estimating disease extent and severity. Radiomics and machine learning may further refine subtle disease detection and diagnostic modeling, whereas deep learning provides automated segmentation, standardized quantification, and pattern classification support[68]. However, diagnostic evidence remains uneven across CTD-ILD subtypes and study tasks. Reported endpoints include ILD detection, subclinical abnormality identification, pattern classification, disease-extent assessment, and evaluation of functional relevance, and the performance reported for one diagnostic task should not be directly extrapolated to another. Broader methodological limitations and priorities for future validation are discussed in the “challenges and future perspectives” section.

HRCT-BASED PROGNOSTIC ASSESSMENT IN CTD-ILD

Beyond the baseline diagnosis, HRCT plays an important role in the prognostic assessment of CTD-ILD. Imaging-derived prognostic information is clinically useful because it helps identify patients at risk of disease progression, pulmonary function decline, progressive pulmonary fibrosis, and mortality. Therefore, prognostic assessment should be viewed as an outcome-oriented process rather than a simple description of baseline CT abnormalities. Visual interpretation and semiquantitative scoring remain the most interpretable tools for baseline risk stratification, whereas quantitative CT, radiomics, machine learning, and deep learning-based analysis have increasingly been investigated for more objective quantification and individualized risk estimation of structural worsening and clinical outcomes[48,69-71].

Visual interpretation and semiquantitative scoring

Visual interpretation and semiquantitative HRCT scoring provide the most established prognostic evidence in CTD-ILD. Their main advantage is clinical interpretability. Disease extent, fibrotic features, inflammatory burden, and dominant HRCT pattern provide clinically interpretable information for baseline risk stratification and may inform multidisciplinary clinical assessment.

Baseline disease extent and physiologic staging: The prognostic value of visual disease extent is best established in SSc-ILD. Goh et al[69] analyzed data of 215 patients with SSc and demonstrated that a visually estimated HRCT disease extent threshold of 20% separated limited from extensive disease. An FVC% predicted threshold of 70% was used as an additional discriminator, forming the classic combined HRCT-PFT staging system, for patients with an indeterminate HRCT extent of 10%-30%. Moore et al[70] subsequently confirmed in a multicenter SSc-ILD cohort that greater baseline HRCT extent predicted both functional decline and mortality. These studies remain important because they provide simple and clinically interpretable thresholds that can be applied in routine risk stratification.

HRCT pattern and early disease progression: Baseline HRCT pattern and semiquantitative extent are useful for identifying patients at risk for early disease progression. This is particularly evident in inflammatory myopathy-associated ILD, where the radiologic pattern frequently reflects subsequent disease behavior. Wu et al[72] retrospectively evaluated 47 patients with antisynthetase syndrome-associated ILD using 211 CT examinations over a median follow-up of 79 months. Radiographic progression occurred in 53.2% of patients and was associated with reduced survival. Among patients with initially non-fibrotic patterns, 56.7% demonstrated progression during follow-up, whereas established fibrotic patterns tended to persist. Tanizawa et al[73] demonstrated that HRCT patterns at diagnosis were prognostically informative in PM/DM-ILD and were associated with anti-MDA5 antibody positivity, thereby further supporting the association between baseline imaging phenotype and disease course.

Semiquantitative HRCT scoring may further improve early progression prediction. Jiang et al[74] retrospectively studied data of 282 patients with PM/DM, including 140 patients with PM/DM-ILD who underwent longitudinal follow-up. During a median follow-up of 5.69 months, 56 patients developed acute or subacute progression. The HRCT score was an independent predictor of progression, and an integrated model combining clinical variables and imaging pattern achieved a C-index of 0.764, with improved reclassification performance. These results indicate that baseline HRCT assessment not only describes disease severity but also helps identify patients at risk of early clinical deterioration.

Semiquantitative scoring has been used to evaluate rapidly progressive ILD risk in anti-MDA5-positive DM-ILD. Zhang et al[11] compared seven HRCT visual scoring methods in 67 patients with anti-MDA5-positive DM. The best performing method divided both lungs into six regions and assigned scores according to the extent of reticulation, honeycombing, and GGO. This method demonstrated the strongest correlation with pulmonary function and achieved an AUC of 0.766 for identifying rapidly progressive ILD. A nomogram integrating clinical and imaging predictors further improved risk prediction, with an internal validation AUC of 0.842.

Inflammatory burden and short-term mortality: Visually assessed inflammatory abnormalities on HRCT are closely associated with short-term mortality in some CTD-ILD subtypes, especially anti-MDA5-positive DM-ILD. Xu et al[75] developed a GGO and consolidation-weighted visual CT model in the derivation and validation cohorts of patients with anti-MDA5-positive DM-ILD. The resulting MDA5 score achieved C-index values of 0.80 and 0.84, with AUCs of 0.85 and 0.87 for 6-month mortality prediction, respectively. Patients above the cutoff of 18 points had markedly worse survival than those below the cutoff. The FLAIR model proposed by Lian et al[76] further integrated ferritin, lactate dehydrogenase, anti-MDA5 antibody level, rapidly progressive ILD status, and CT imaging score in 207 patients with amyopathic DM-ILD, separating patients into low, medium, and high-risk groups with substantially different 1-year survival rates. Further, Liu et al[77] reported that higher GGO scores were associated with stepwise increases in mortality in anti-MDA5-positive DM. These studies indicate that HRCT-based assessment of inflammatory burden may be particularly valuable for short-term risk stratification in rapidly progressive CTD-ILD phenotypes.

Quantitative CT

Quantitative CT refines prognostic assessment by converting structural abnormalities into continuous numerical measures. Quantitative metrics may provide more reproducible estimates of fibrotic burden, texture abnormality, lung volume, vascular features, and radiological progression compared with broad visual categories.

Baseline quantitative burden, severity, and survival risk: Quantitative CT can provide objective markers of disease severity and survival risk. Bertolazzi et al[78] found that quantitative CT parameters correlated moderately with semiquantitative CT scores and differed significantly between severe and mild ILD in RA-ILD. These parameters also demonstrated good discriminative ability for severity stratification, with AUCs ranging from 0.73 to 0.81. This study focused primarily on severity assessment; however, it supports further evaluation of quantitative CT as an adjunct to semiquantitative scoring.

Quantitative fibrotic burden has a clearer prognostic value. Oh et al[79] retrospectively analyzed 144 patients with RA-ILD, using an automated quantification system and reported a 5-year mortality of 30.5% after a median follow-up of 52.2 months. Quantitative lung fibrosis, defined as the sum of reticulation and traction bronchiectasis, exhibited the strongest prognostic value. Patients with quantitative lung fibrosis of ≥ 12% of total lung volume had a 5-year mortality of 50.0%, compared with 17.4% among those below this threshold. Amorim et al[80] found that CT-derived reticular burden was also associated with survival in SSc-ILD. A baseline reticular pattern of ≥ 1.41% and a follow-up reticular burden of ≥ 4.34% were both associated with worse outcomes. These studies indicate that quantitative measures of fibrosis and reticulation can provide clinically meaningful risk stratification before more complex modeling is applied.

Radiological progression and progressive pulmonary fibrosis: Quantitative CT may help define radiological progression and progressive pulmonary fibrosis. Beck et al[81] included 91 patients with CTD-ILD and showed a ≥ 2.2% increase in CT-derived fibrosis volume over 2 years as the optimal threshold for defining radiological progression. Patients who met a combined definition of progressive pulmonary fibrosis based on both CT progression and pulmonary function deterioration demonstrated significantly higher mortality, and this combined definition was the only independent predictor of mortality in multivariable analysis. This study is important because it links quantitative structural progression to a clinically relevant outcome, thereby supporting the integration of CT-derived fibrosis change with pulmonary function criteria.

Vascular, histogram-based, and integrated quantitative biomarkers: Beyond parenchymal fibrosis, quantitative CT can capture vascular and histogram-based features that may have prognostic value. Qiang et al[82] conducted a two-center retrospective study of 578 patients with IIM-ILD and conducted AI-based quantitative analysis of baseline HRCT to evaluate pulmonary vessel-related structure and interstitial abnormality parameters. Patients with rapidly progressive ILD demonstrated higher mean pulmonary vessel diameter, pulmonary vessel-related structure volume, and standard deviation of pulmonary vessel diameter than those without rapidly progressive disease. Subtype-specific models in antisynthetase syndrome and anti-MDA5-positive DM exhibited good discrimination and were externally validated, indicating that vascular-related CT parameters may help identify patients at risk for rapid progression and poor prognosis.

Histogram-based analysis provides prognostic information through pattern classification. Chikasue et al[57] developed a whole-lung CT volume histogram analysis model to differentiate UIP from NSIP. The model achieved an AUC of 0.91 in the training cohort and 0.81 in the external validation cohort for identifying a UIP pattern. Importantly, the Veterans Health Administration-derived score was significantly associated with prognosis in multivariable analysis. This evidence comes from fibrosing interstitial pneumonia cohorts rather than a dedicated CTD-ILD population; however, it supports the concept that quantitative pattern-related features may carry prognostic information.

AI-assisted quantitative HRCT may provide a broader assessment of severity and survival risk. Chu et al[35] extracted 17 imaging parameters from HRCT in 116 patients with CTD-ILD. These parameters differed across pulmonary function-based severity groups, and several imaging metrics were associated with survival. Left lung volume and the percentage of lesional components of ≤ -751 HU remained independent prognostic factors in multivariable Cox regression. These findings indicate that integrated quantitative HRCT analysis may capture both the extent and density distribution of lung involvement, despite requiring larger external validation studies.

Radiomics and machine learning

Radiomics and machine learning extend prognostic assessment by capturing high-dimensional imaging heterogeneity beyond routine visual review and conventional quantitative measurements. Their potential value lies in the individualized prediction of progression, mortality, and survival risk, particularly when integrating imaging features with clinical and functional variables.

Radiomic risk scores for progression and staging: Schniering et al[71] developed and externally validated a quantitative radiomic risk score in the derivation and validation cohorts of 90 and 66 patients with SSc-ILD, respectively. Patients classified as high risk had substantially shorter progression-free survival, and high-risk status was related to a hazard ratio of 5.14 for progression. This study provides one of the clearest examples of radiomics being used not only to describe HRCT heterogeneity but also to stratify progression risk.

Radiomics has been explored for staging-based risk assessment in mixed CTD-ILD. Qin et al[36] developed a non-contrast CT-based radiomics nomogram for ILD-gender, age, and physiology (GAP) staging in 245 patients from two centers. The model achieved AUCs of 0.887, 0.885, and 0.850 in the training, internal validation, and external validation cohorts, respectively. ILD-GAP staging is not a direct survival endpoint; however, this work indicates that radiomics may provide imaging-based support for established physiologic risk stratification systems.

Mortality prediction in high-risk CTD-ILD populations: The prognostic value of HRCT radiomics is particularly well illustrated in anti-MDA5-positive DM-ILD. Xu et al[83] developed a CT radiomics-based model for 6-month mortality in 228 patients from two centers. Their final Rad-score plus model, integrating radiomics with age and FVC% predicted, achieved C-index values of 0.88, 0.88, 0.83, and 0.84 across the training, testing, internal validation, and external validation datasets, respectively, outperforming visual CT scoring and ILD-GAP. Further, He et al[84] developed a multicenter clinico-radiologic-radiomic nomogram for predicting rapidly progressive ILD and mortality in anti-MDA5-positive DM. The bilateral lung radiomics score achieved AUCs of 0.898, 0.869, and 0.905 in the training, internal validation, and external validation datasets, respectively, and the final integrated nomogram further improved performance. Moreover, high-risk patients had significantly shorter survival. These studies indicate that anti-MDA5-positive DM-ILD is currently one of the most developed settings for prognostic HRCT radiomics in CTD-ILD.

Radiomics-based survival prediction has been studied in RA-ILD. Liu et al[85] included 230 patients with RA-ILD, divided into training, internal validation, and external validation cohorts. From 1688 radiomic features, 24 were selected by LASSO-Cox regression to construct a Rad-score. The integrated model combining Rad score with clinical variables achieved C-indices of 0.832, 0.816, and 0.812 in the three cohorts, respectively. Patients in the high-risk group demonstrated significantly shorter overall survival than those in the low-risk group. This study further demonstrates the potential of radiomics-based prognostic modeling in RA-ILD, although validation in larger and more diverse cohorts is still needed.

Integrated clinical imaging and machine learning models: Machine-learning models may further improve prognostic discrimination by integrating imaging features with clinical, physiologic, and serologic variables. Xu et al[86] conducted a multicenter retrospective study of patients with CTD-ILD diagnosed through multidisciplinary assessment and compared three prognostic systems: ILD-GAP and fibrosis score as independent metrics, a composite model combining GAP index and fibrosis score, and a machine-learning model incorporating clinical and imaging variables. The fibrosis score performed better than ILD-GAP alone; the composite model further improved discrimination; and random forest and support vector machine classifiers achieved the best performance among machine learning approaches. Shapley additive explanations analysis identified fibrotic extent, DLCO, and age as the main contributors to adverse outcome prediction. These results reinforce the prognostic importance of fibrotic burden while suggesting additional value from integrating imaging findings with established clinical and functional predictors through machine learning.

Deep learning and longitudinal imaging analysis

Deep learning and longitudinal imaging analysis move HRCT-based prognosis from static baseline risk estimation toward dynamic recognition of structural worsening. These methods may be particularly useful for detecting subtle interval changes, quantifying pattern evolution, and linking imaging trajectories to pulmonary function decline or survival.

Deep learning-based structural change and quantitative prognostic biomarkers: Deep learning-based HRCT analysis may provide prognostic information from both serial imaging change and baseline quantitative outputs. Chassagnon et al[87] applied elastic registration-driven deep learning to serial CT scans in SSc. The method identified morphologic worsening with an accuracy of 80% and functional worsening with an accuracy of 83% in an independent test set of 40 patients. Further, deformation-derived lung shrinkage metrics substantially correlated with changes in FVC and DLCO. These results suggest that serial CT-based deformation analysis may provide a useful means of identifying structural worsening associated with functional decline.

Deep learning outputs derived from baseline HRCT may also function as prognostic biomarkers. Stock et al[88] examined a deep learning HRCT algorithm for UIP probability in SSc-ILD and revealed that higher algorithm-derived UIP probability was associated with more advanced disease, greater progression, and worse survival. Thus, an algorithmic output originally developed for pattern recognition may also provide prognostic information.

Deep learning-based quantitative analysis of baseline HRCT has also been used to assess rapid progression and adverse outcomes in IIM-ILD. Greater GGO and consolidation burden were associated with rapidly progressive disease in antisynthetase syndrome-associated ILD in a cohort of 511 patients with IIM-ILD. Among patients with anti-MDA5-positive dermatomyositis-associated ILD, and a history of rapidly progressive ILD, quantitative GGO measures were associated with adverse events, whereas neither GGO nor consolidation was associated with adverse events in those without such a history[89]. These findings indicate that baseline deep learning-derived inflammatory burden may have prognostic value, particularly in myositis-associated ILD.

Longitudinal imaging trajectory and time-dependent risk: Longitudinal deep learning analysis further characterizes how CT abnormalities evolve. Qiang et al[90] enrolled 514 patients with IIM-ILD and conducted quantitative analysis of interstitial abnormalities on HRCT using a deep learning algorithm. Anti-MDA5 antibody positivity, reduced FVC, and greater baseline GGO and reticulation burden were independent risk factors for rapidly progressive ILD. Changes in reticulation and consolidation correlated with changes in pulmonary function, and the prognostic impact of GGO proportion on adverse outcomes increased over time. These findings indicate that HRCT evolution trajectories differ across clinical subtypes and that longitudinal GGO may serve as a time-dependent prognostic indicator.

Summary and current evidence gaps

Overall, HRCT-based prognostic assessment in CTD-ILD should be interpreted as a layered and outcome-oriented framework. Visual and semiquantitative methods provide immediately interpretable baseline risk stratification, particularly through disease extent, HRCT pattern, and inflammatory burden. Quantitative CT captures continuous measures of fibrotic burden, texture abnormality, lung volume, vascular-related structure, and radiological progression. Radiomics and machine learning refine individualized prediction of progression and mortality, whereas deep learning-based longitudinal analysis provides a route toward earlier detection of clinically important structural worsening[69,71,79,87].

However, prognostic evidence remains heterogeneous across CTD-ILD subtypes and outcome definitions. Functional decline, radiological progression, progressive pulmonary fibrosis, short-term mortality, and overall survival represent distinct endpoints and should not be considered interchangeable when interpreting or comparing model performance. Broader issues related to cohort heterogeneity, external validation, and clinical translation are discussed in the “challenges and future perspectives” section.

HRCT-BASED LONGITUDINAL MONITORING AND TREATMENT RESPONSE ASSESSMENT IN CTD-ILD

Longitudinal monitoring in CTD-ILD relies on the integration of symptoms, PFTs, serological markers, treatment exposure, and serial HRCT findings. Follow-up HRCT is used to identify whether lung abnormalities improve, remain stable, or progress over time, including changes in GGO, consolidation, reticulation, traction bronchiectasis, honeycombing, and overall fibrotic extent. However, subtle interval changes are frequently challenging to judge consistently through visual review alone, especially in patients receiving long-term immunomodulatory or antifibrotic therapy. Therefore, HRCT-based monitoring should be considered a layered process that includes visual interpretation, semiquantitative scoring, quantitative CT, AI-assisted analysis, and longitudinal pattern-transition assessment. Current evidence is strongest in SSc-ILD and treatment trial cohorts, whereas data in broader CTD-ILD populations remain more limited[41,91,92].

Visual interpretation and semiquantitative follow-up

Visual interpretation and semiquantitative scoring remain the most clinically accessible approaches for follow-up assessment. Their main value lies in providing an interpretable description of whether HRCT abnormalities have improved, stabilized, or worsened during treatment. These approaches are less sensitive than quantitative tools for small interval changes; however, they remain important reference methods in clinical practice and therapeutic studies.

Serial HRCT assessment of structural improvement and stabilization: Serial HRCT scoring has been used to assess structural improvement after treatment, particularly in SSc-ILD. Launay et al[92] used quantitative HRCT scoring to longitudinally assess pulmonary changes after autologous hematopoietic stem cell transplantation in patients with systemic sclerosis. During a median follow-up of 60 months, the overall HRCT disease-extent score decreased from 10 at baseline to 4 at 6 months after transplantation (P = 0.04). However, this early radiological improvement was not sustained throughout follow-up, as the score subsequently increased before reaching relative stability. Serial HRCT can therefore provide imaging evidence of structural response when treatment produces substantial disease regression.

Moreover, visual and semiquantitative follow-up may help identify patients who remain at risk of radiological progression despite therapy. Cosău et al[93] conducted a real-world study of 20 patients with SSc-ILD treated with mycophenolate mofetil. Pulmonary function remained largely stable during treatment; however, patients with UIP patterns had higher baseline Warrick scores and greater radiological progression than those with an NSIP pattern. Further, the Warrick score correlated moderately with DLCO, and patients who later received add-on nintedanib had more extensive radiological involvement and lower baseline DLCO. Thus, follow-up HRCT may help document disease stability while identifying patients with persistent structural risk who may require treatment escalation.

HRCT pattern as an indicator of treatment response: Baseline HRCT pattern may help explain differences in treatment response across CTD-ILD subtypes. A recent preprint by Bolig et al[94] evaluated 41 patients with myositis-associated ILD with circulating myositis-associated antibodies who received immunomodulatory therapy. Patients with an OP-predominant pattern demonstrated significant improvement in FVC, DLCO, quantitative GGO or consolidation scores, and qualitative radiological assessments at 24 months. Conversely, patients with an NSIP-predominant pattern showed no significant pulmonary function improvement. These findings suggest that OP-predominant and NSIP-predominant patterns may exhibit different treatment-associated functional and radiologic trajectories; however, the small retrospective cohort and non-peer-reviewed status of the study preclude causal inference regarding differential treatment efficacy.

Similar observations have been reported in smaller treatment cohorts. Okamoto et al[95] described six patients with acute exacerbation of RA-associated ILD treated with Janus kinase inhibitors and systemic corticosteroids. Respiratory status improved in all patients, and CT scores decreased post-treatment, particularly in the fibrotic component, although the small sample size limits interpretation. Upadacitinib treatment in refractory IIM-ILD was associated with significant improvement in FVC and FVC% predicted, whereas HRCT Warrick scores remained largely stable overall, with some patients exhibiting radiological improvement[96]. Further, Zhu et al[97] reported that baricitinib treatment in anti-MDA5-positive DM-ILD was associated with a significant decrease in HRCT score and improvement in FVC. Observed imaging improvements should therefore be interpreted as treatment-associated changes rather than definitive evidence of treatment efficacy, particularly in uncontrolled cohorts.

Quantitative CT for serial monitoring

Quantitative CT adds an objective dimension to longitudinal monitoring by measuring interval changes in lung density, lung volume, parenchymal texture, and disease extent. Compared with visual review, quantitative approaches may be better suited for detecting small structural changes and linking imaging progression to functional decline.

Quantitative change, functional correlation, and visual agreement: Quantitative CT metrics can be compared with visual scores and pulmonary function to identify whether structural changes are clinically meaningful. A prospective study of 57 patients with fibrosing ILD, most commonly CTD-ILD, conducted Computer-Aided Lung Informatics for Pathology Evaluation and Rating (CALIPER) analysis, Warrick semiquantitative scoring, and PFT at baseline and 6 months. CALIPER GGO and fibrosis scores correlated with corresponding Warrick indices, and the CALIPER fibrosis score at 6 months demonstrated a strong correlation with desaturation during the 6-minute walk test. A 5% reduction in normal lung volume exhibited moderate accuracy for predicting disease progression[98]. These findings suggest that quantitative CT may complement semiquantitative scoring and functional assessment during follow-up.

Longitudinal quantitative CT has been evaluated in broader CTD-ILD cohorts. Machado et al[99] studied 195 patients with CTD-ILD who underwent quantitative and visual HRCT assessment at baseline and two follow-up visits. Lung volume markedly decreased over time, and the fibrosis index increased by the second follow-up. Decreases in lung volume and increases in disease extent were associated with FVC decline, and these associations remained consistent across SSc and non-SSc subgroups. Quantitative CT may therefore complement PFT by providing structural markers of longitudinal worsening.

Quantitative HRCT in treatment monitoring: Quantitative CT has been used to assess the association between treatment and stabilization or improvement of structural lung disease. Occhipinti et al[100] evaluated 35 patients with SSc-ILD who underwent CT and PFT before and after immunosuppressive therapy. Changes in individual texture patterns had limited ability to predict functional progression; however, quantitative change in total lung volume predicted a composite functional endpoint based on FVC% predicted and DLCO% predicted, with an AUC of 0.74. This indicates that lung volume change may be a useful global marker in longitudinal assessment.

Antifibrotic treatment studies further support the role of quantitative HRCT for monitoring stabilization. Di Battista et al[101] followed 10 patients with SSc-ILD treated with nintedanib for 12 months and revealed no significant change in CALIPER-derived percentages of normal parenchyma or ILD components, paralleling stabilization on PFT. A subsequent prospective study of patients with fibrosing ILD, including SSc-ILD and IPF, revealed that nintedanib-treated patients demonstrated relative stability in normal lung parenchyma and ILD volume, whereas untreated historical controls showed loss of normal lung and progression of ILD volume[102]. Together, these observations support quantitative HRCT for documenting treatment-associated stabilization, particularly in antifibrotic therapy cohorts.

Quantitative baseline features may help identify patients more likely to benefit from specific treatments. Matson et al[103] studied patients with SSc-ILD and RA-ILD and revealed that higher quantitative fibrosis and higher GGO were associated with FVC% predicted improvement in patients with SSc-ILD treated with cyclophosphamide. No similar associations were observed in the mycophenolate mofetil-treated subgroup or in RA-ILD patients. This finding requires further validation but suggest that baseline quantitative HRCT features may help identify imaging phenotypes associated with differential treatment-related functional change.

Quantitative HRCT in trial cohorts and structured follow-up: The strongest evidence for HRCT-based monitoring of treatment-related structural change comes from transplantation cohorts and randomized trials in SSc-ILD. Wada et al[104] assessed 33 patients with SSc treated with autologous hematopoietic stem cell transplantation and found that patients with improved pulmonary function demonstrated significant reductions in pulmonary densities, supporting quantitative HRCT density analysis as a biomarker of treatment response. Pugnet et al[105] further reported significant regression of ILD and GGO extent at 12 months and 24 months post-transplantation in diffuse SSc. Further, radiological responders demonstrated a trend toward better 5-year survival than non-responders.

Randomized trial data provide additional support. In Scleroderma Lung Study II, Goldin et al[91] analyzed volumetric HRCT scans from 97 participants treated with cyclophosphamide or mycophenolate mofetil. QILD, defined as the sum of quantitative fibrosis, GGO, and honeycombing, significantly decreased in the pooled treatment groups. Further, longitudinal change in whole-lung QILD was correlated with changes in FVC, DLCO, and dyspnea score. These findings are important because they associate HRCT-derived structural change with functional and symptomatic improvement in a controlled treatment setting.

Quantitative densitometry and autoimmune-featured ILD: Quantitative densitometry may help monitor disease stage and progression in RA-ILD. Guo et al[106] evaluated 138 patients with RA-ILD stratified by GAP stage and measured the percentage of normal lung attenuation, low-attenuation area, and high-attenuation area in each lobe. With the advancing ILD stage, normal lung attenuation decreased, whereas low and high-attenuation areas increased. The combination of these densitometric parameters exhibited moderate performance for early diagnosis and progression monitoring. This study was not primarily treatment-response based; however, it supports the use of quantitative CT features for staging and monitoring RA-ILD progression.

Related evidence is available in interstitial pneumonia with autoimmune features (IPAF). Pušeljić et al[107] compared quantitative CT features among CTD-ILD, IPAF, and IPF and found that IPAF exhibited intermediate progression-free survival between CTD-ILD and IPF. A higher GGO/consolidation ratio and emphysema percentage were associated with a lower risk of progression. Further, Zhao et al[108] reported that in patients with IPAF treated with tacrolimus, the total HRCT score significantly decreased post-treatment, with marked improvement in consolidation and GGO components. Baseline consolidation score independently predicted radiological improvement. These studies indicate that quantitative HRCT may be useful for monitoring autoimmune-featured ILD, although IPAF should be interpreted as a related but not identical population to defined CTD-ILD.

AI-assisted quantification in longitudinal assessment

AI-assisted quantification may improve consistency across readers and provide a more reproducible assessment of interval change. This is particularly relevant in longitudinal monitoring, where small differences in visual estimation may affect judgments of treatment response or disease progression.

A study of 48 patients with CTD-ILD receiving antifibrotic therapy performed four HRCT examinations in each patient. A senior chest radiologist and a non-senior reader independently assessed abnormal lung patterns using semiquantitative scoring, whereas an AI-based lung texture analysis platform generated continuous volumetric percentages for hyperlucency, GGO, reticulation, and honeycombing. The AI system demonstrated substantial agreement with the senior radiologist, with an overall concordance rate of 81% and low mean absolute error. The agreement between the AI system and the non-senior reader was significantly lower[41]. In this cohort, AI-derived quantitative measurements showed substantial agreement with assessment by an experienced radiologist, suggesting potential to support more standardize serial HRCT assessment across readers with different expertise levels.

Deep learning-radiomics has been explored for imaging-based severity stratification in CTD-ILD. Long et al[109] extracted deep learning-radiomics features from CT images of 264 patients with CTD-ILD and developed models for predicting the GAP stage. The combined model incorporating clinical variables and deep learning-radiomics features achieved the highest performance, with an AUC of 0.951. This study focused on the GAP stage rather than direct serial change; however, it supports the broader concept that AI-derived CT features can be integrated with clinical variables for risk stratification during longitudinal management.

Longitudinal pattern-transition and reversibility analysis

Longitudinal pattern-transition analysis provides a more granular approach to monitoring treatment response. Instead of measuring only total disease extent, this approach assesses whether abnormal lung patterns convert toward more normal parenchyma, remain stable, or progress toward more fibrotic abnormalities.

In Scleroderma Lung Study I, Kim et al[110] analyzed paired baseline and 12-month HRCT scans from 83 participants treated with oral cyclophosphamide or placebo. Patients receiving cyclophosphamide demonstrated more favorable transitions from fibrotic or GGO patterns to normal lung, whereas placebo-treated patients showed changes in the opposite direction. This approach was further extended in Scleroderma Lung Study II using volumetric baseline and 24-month HRCT scans from patients treated with cyclophosphamide or mycophenolate mofetil. Both treatment groups demonstrated favorable transitions from GGO or fibrosis to normal lung[111]. These studies indicate that treatment response may involve not only a reduction in total disease burden but also conversion of specific abnormal imaging patterns.

Pattern reversibility has been studied in broader CTD-ILD populations. Zhang et al[112] developed a standardized protocol for assessing lesion reversibility in CTD-ILD and constructed a nomogram integrating radiomic and visual features to predict baseline reversibility. The study included 153 patients and 575 involved pulmonary zones. Lesions were classified as completely or non-completely reversible according to longitudinal CT follow-up. The combined radiomic-visual nomogram achieved the best performance, with AUCs of 0.86, 0.90, and 0.82 in the training, internal validation, and external validation cohorts, respectively. This study provides a useful example of moving from patient-level progression assessment toward lesion-level reversibility prediction.

Deep learning-based quantification may further support treatment-response assessment in high-risk inflammatory CTD-ILD. You et al[113] evaluated 70 patients with anti-MDA5-positive DM, including 39 treated with tofacitinib and 31 not receiving tofacitinib. Deep learning-based HRCT analysis revealed a significant reduction in whole-lung involvement in the tofacitinib group, with a greater reduction in total lesion volume compared with the non-tofacitinib group. Further, the tofacitinib group had a higher 3-year survival rate, and tofacitinib treatment was identified as an independent protective factor against mortality. These findings indicate that AI-based quantitative HRCT may provide an objective way to assess treatment-associated lesion regression.

Longitudinal growth-rate modeling provides another approach to capturing heterogeneous disease courses. Tian et al[114] analyzed data of 80 patients with IIM-ILD and 243 HRCT scans using quantitative lung fibrosis scores and a growth-rate model. Five longitudinal patterns were identified: Progressive, improving, convex, concave, others and stable. A rapid-progression subgroup defined by the median progression rate was significantly associated with mortality among patients with progressive disease. This approach indicates that longitudinal quantitative modeling may help identify clinically relevant subgroups from real-world serial CT data, although further validation is necessary.

Summary and current evidence gaps

Overall, HRCT-based longitudinal monitoring in CTD-ILD should be considered a continuum rather than a single measurement strategy. Visual and semiquantitative follow-up remain clinically accessible and interpretable, quantitative CT provides more objective assessment of small interval changes, AI-assisted tools may improve consistency across readers, and pattern-transition analysis provides insight into whether therapy is associated with structural improvement, stabilization, or progression[41,91,111].

However, the strength of longitudinal evidence remains variable across CTD-ILD subtypes and therapeutic settings. In particular, imaging improvement or stabilization observed during therapy should generally be interpreted as treatment-associated structural change rather than evidence of treatment efficacy unless supported by appropriately controlled prospective data. Broader methodological limitations and priorities for standardization are discussed in the “challenges and future perspectives” section.

CHALLENGES AND FUTURE PERSPECTIVES

Despite substantial progress, several challenges continue to limit the clinical translation of HRCT-based assessment in CTD-ILD. Visual interpretation and semiquantitative scoring remain practical and widely applied; however, reader experience, interobserver variability, and differences in the definition of subtle GGO, early reticulation, traction bronchiectasis, and mixed inflammatory-fibrotic patterns influence their reproducibility. This limitation is particularly relevant during longitudinal follow-up when small interval changes may influence treatment decisions. Quantitative CT, radiomics, machine learning, and deep learning features may reduce some aspects of subjective interpretation; however, they introduce their own sources of variability, including CT acquisition parameters, reconstruction algorithms, inspiratory effort, segmentation strategy, feature extraction workflow, and postprocessing software. Without standardized protocols, the reproducibility and comparability of HRCT-derived imaging biomarkers across centers remain limited. Recent guidance for medical imaging AI, including the updated Checklist for AI in Medical Imaging and the FUTURE-AI international consensus guideline for trustworthy and deployable AI in healthcare, further emphasizes transparent and reproducible reporting, robust validation, and consideration of real-world clinical deployment[115,116].

Although quantitative CT has emerged as an objective adjunct in selected clinical and research settings, its implementation remains variable across centers. By comparison, radiomics, machine learning, and deep learning have not yet achieved sufficient validation for routine clinical implementation in CTD-ILD. One major challenge is the relatively low reproducibility of quantitative and computational imaging features across different scanners, acquisition protocols, reconstruction methods, and analysis platforms. In addition, some approaches require specialized software, computational resources, and additional maintenance costs, which may limit their availability in routine practice. Furthermore, the clinical applicability of advanced HRCT-based approaches may vary across healthcare systems with different levels of imaging infrastructure, computational resources, and specialized expertise. Future studies should therefore evaluate not only technical performance but also feasibility and clinical utility across diverse healthcare environments. Integration into existing radiology workflows can also be challenging, particularly when additional image processing or manual interaction is required. Radiation exposure is another practical consideration, particularly when serial HRCT examinations are required for longitudinal monitoring. Low-dose and reduced-acquisition protocols may help reduce cumulative radiation burden, but their ability to preserve the reproducibility of quantitative and computational imaging biomarkers should be carefully validated before widespread implementation[39,50,60]. Therefore, wider clinical implementation will require more standardized and reproducible workflows, accessible software platforms, and reduced implementation and maintenance costs.

Taken together, current evidence suggests a gradient of clinical readiness rather than equivalent applicability across HRCT-based approaches. Visual interpretation and semiquantitative assessment are sufficiently established to inform routine clinical evaluation, whereas quantitative CT is better considered an objective adjunct that may be useful in appropriately equipped centers or structured research and trial settings. Radiomics, machine learning, and deep learning should currently remain primarily research tools until reproducibility, calibration, external validation, workflow integration, and clinical utility are demonstrated prospectively. Their reported diagnostic or prognostic performance should therefore not be interpreted as evidence of immediate readiness for routine clinical deployment.

The biological and clinical heterogeneity of CTD-ILD further complicates model development and validation. SSc-, RA-, inflammatory myopathy-, Sjögren syndrome-, and SLE-associated ILD differ in terms of dominant HRCT patterns, progression rates, treatment response, and prognostic determinants. Therefore, models developed in one CTD subtype may not be directly applicable to another. Further, current evidence is unevenly distributed across CTD-ILD subtypes. Relatively more data are available for SSc-, RA-, and inflammatory myopathy-associated ILD, particularly anti-MDA5-positive DM-associated ILD, whereas studies on Sjögren syndrome- and SLE-associated ILD remain comparatively limited. Although RA-ILD has been increasingly investigated using semiquantitative assessment, quantitative CT, radiomics, and prognostic modeling, the distribution of evidence across diagnostic, prognostic, and longitudinal monitoring applications remains variable. In contrast, advanced HRCT-based studies in Sjögren syndrome-associated ILD and SLE-associated ILD remain relatively sparse. Mixed CTD-ILD cohorts broaden the range of studied populations but may introduce additional heterogeneity because of differences in disease subtype, HRCT phenotype, treatment exposure, and outcome definitions, thereby limiting subtype-specific interpretation and model generalizability. Further, several studies are retrospective, single-center, and based on small or selected populations, and many proposed models still lack robust multicenter external validation. Therefore, differences in reported diagnostic or prognostic performance across studies should be interpreted in the context of variations in CTD subtype, cohort composition, imaging acquisition and analysis protocols, outcome definitions, and validation strategies, rather than being regarded as direct evidence of superiority of one HRCT-based approach over another. These methodological limitations increase susceptibility to selection bias and model overfitting and may lead to optimistic estimates of diagnostic or prognostic performance, thereby limiting the generalizability of reported findings to broader real-world CTD-ILD populations. Interpretation of longitudinal and treatment-response studies requires additional caution because observed imaging changes may also be influenced by baseline disease severity, natural disease trajectory, concomitant therapies, and treatment-selection bias. Accordingly, associations between HRCT improvement or stabilization and treatment exposure should not be interpreted as evidence of treatment efficacy unless supported by appropriately controlled prospective studies.

Another unresolved issue is the lack of harmonized outcome definitions. Existing studies have examined different tasks, including prevalent ILD detection, baseline disease-extent assessment, physiologic impairment, radiologic progression, treatment response, short-term mortality, and long-term survival. These endpoints are not interchangeable. Similarly, progression has been characterized by HRCT worsening, FVC decline, DLCO decline, symptom deterioration, or composite criteria. Such heterogeneity makes it challenging to compare models, identify clinically meaningful thresholds, and translate imaging biomarkers into practice. Therefore, future studies should clearly distinguish diagnostic, prognostic, and monitoring tasks and predefine clinically relevant endpoints, time horizons, and response criteria. Taken together, these methodological constraints indicate that the reported performance of individual HRCT-based approaches should be interpreted cautiously and should not be assumed to generalize across CTD subtypes, institutions, or clinical settings without further prospective external validation.

Future work should prioritize standardized HRCT acquisition and analysis protocols, transparent segmentation and feature extraction workflows, and prospective multicenter validation across well-characterized CTD-ILD subtypes. Moreover, greater attention should be paid to model interpretability, calibration, decision curve analysis, and clinical utility rather than discrimination metrics alone. Multimodal integration represents an important future research direction. HRCT-derived metrics may be combined with clinical variables, autoantibody profiles, inflammatory biomarkers, pulmonary function, treatment exposure, and longitudinal follow-up data to develop more comprehensive models. Future studies should determine whether such multimodal approaches provide reproducible incremental value over established clinical and radiologic assessment and whether they can improve clinically meaningful decision-making. Ultimately, the goal is not to replace expert radiologic interpretation but to develop standardized and interpretable imaging biomarkers that can support earlier diagnosis, individualized risk stratification, and objective monitoring of treatment response in CTD-ILD. The clinical readiness of current HRCT-based approaches and the conceptual progression toward multimodal integration, together with key priorities for future validation and clinical translation, are summarized in Figure 1.

Figure 1 Clinical readiness and future directions of high-resolution computed tomography-based assessment in connective tissue disease-associated interstitial lung disease. Visual interpretation and semiquantitative scoring represent established clinical approaches, whereas quantitative computed tomography serves as an emerging objective adjunct with variable implementation across centers. Radiomics, machine learning, and deep learning remain predominantly investigational. Multimodal integration represents a conceptual future research direction requiring prospective validation, standardized and reproducible workflows, harmonized endpoints, interpretability, and demonstration of clinical utility before broader implementation. This original schematic was designed and constructed by the authors using Microsoft PowerPoint. The graphical icons were selected from the open-source Tabler Icons library and used under the MIT License. HRCT: High-resolution computed tomography; CTD-ILD: Connective tissue disease-associated interstitial lung disease; CT: Computed tomography.

CONCLUSION

HRCT remains central to the diagnosis, prognostic assessment, and longitudinal monitoring of CTD-ILD. Visual interpretation and semiquantitative scoring constitute the established foundation of clinical assessment because of their accessibility, interpretability, and direct relevance to routine practice. Quantitative CT provides a more objective assessment of disease extent and structural change and may serve as an emerging adjunct in selected centers and structured research or trial settings. Radiomics, machine learning, and deep learning offer promising capabilities for high-dimensional phenotyping, automated image analysis, and risk prediction but remain predominantly investigational.

This review integrates conventional and advanced HRCT approaches within a clinically oriented framework that reflects their differing levels of evidence maturity and clinical readiness. Importantly, advanced quantitative and computational biomarkers should not currently replace expert radiologic interpretation or established clinical evaluation. The maturity of evidence also varies substantially across CTD-ILD subtypes: Diagnostic and prognostic studies are relatively more developed in SSc-, RA-, and inflammatory myopathy-associated ILD, particularly anti-MDA5-positive DM-associated ILD, whereas evidence for Sjögren syndrome- and SLE-associated ILD remains comparatively limited. Future work should prioritize standardization, reproducibility, harmonized endpoints, prospective multicenter external validation, model calibration and interpretability, and demonstration of incremental clinical utility before broader implementation. With continued methodological refinement and multimodal validation, advanced HRCT-based approaches may ultimately contribute to more objective and individualized CTD-ILD assessment.