Salman A, Elewa A, Marwan A. Artificial intelligence in autoimmune pancreatitis: Differentiating immunoglobulin G4-related pancreatitis from pancreatic ductal adenocarcinoma. Artif Intell Gastroenterol 2026; 7(3): 123890 [DOI: 10.35712/aig.123890]
Corresponding Author of This Article
Ahmed Salman, Internal Medicine, Kasr Alainy School of Medicine, 1 Al-Saray Street, Cairo 11562, Egypt. awea844@gmail.com
Research Domain of This Article
Gastroenterology & Hepatology
Article-Type of This Article
review-article
Open-Access Policy of This Article
This article is an open-access article which was selected by an in-house editor and fully peer-reviewed by external reviewers. It is distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited and the use is non-commercial. See: http://creativecommons.org/licenses/by-nc/4.0/
Author contributions: Salman A conceived and designed the review, performed the literature search, and drafted the manuscript; Elewa A critically revised the manuscript for important intellectual content and approved the final version; Marwan A conceived the idea of the work and contributed to drafting and critical revision; and all authors read and approved the final version of the manuscript.
AI contribution statement: The authors used artificial-intelligence-based tools (ChatGPT 5.6) solely for language editing and grammatical corrections. No artificial-intelligence tool was used to generate, analyze, or interpret data, to draft scientific content, to formulate conclusions, or to select references. All content was reviewed and verified by the authors, who take full responsibility for the accuracy, originality, and integrity of the manuscript.
Conflict-of-interest statement: There are no conflicts of interest to report.
Corresponding author: Ahmed Salman, Internal Medicine, Kasr Alainy School of Medicine, 1 Al-Saray Street, Cairo 11562, Egypt. awea844@gmail.com
Received: June 2, 2026 Revised: July 30, 2026 Accepted: August 10, 2026 Published online: November 8, 2026 Processing time: 155 Days and 23.8 Hours
Abstract
Autoimmune pancreatitis (AIP), specifically the immunoglobulin G4-related type-1 variant, serves as a significant benign mimic of pancreatic ductal adenocarcinoma (PDAC). Focal or mass-forming AIP can manifest as obstructive jaundice, a pancreatic mass, ductal narrowing, weight loss, and elevated tumor markers that may resemble those of PDAC. Misdiagnosis can result in either unnecessary pancreatic surgery or delayed treatment for cancer. Artificial intelligence (AI), including radiomics, machine learning, deep learning, and multimodal decision-support systems, offers a possible solution for this diagnostic dilemma. AI-based imaging is able to extract quantitative characteristics from a variety of modalities, such as computed tomography, magnetic resonance imaging, diffusion-weighted imaging, positron emission tomography, and endoscopic ultrasonography that may not be apparent by visual inspection. Digital pathology and models combining radiological, serological, histological and clinical variables are also on the rise. Current data is limited by small retrospective single-center cohorts, uneven imaging techniques, dependency on segmentation, and the absence of adequate external validation. AI should be seen as a supplement, not a substitute, to the professional multidisciplinary assessment, boosting the risk classification, directing the tissue collection, and steroid trials, and avoiding unnecessary procedures without delaying PDAC diagnosis. Future research should target prospective multicenter validation, uniform datasets, improved explainability and therapeutically significant objectives.
Core Tip: Autoimmune pancreatitis, especially localized immunoglobulin G4-related condition, is very similar to pancreatic ductal adenocarcinoma and constitutes a great diagnostic dilemma. This review summarizes the emerging applications of artificial intelligence (AI) such as radiomics, deep learning, endoscopic ultrasound image analysis, digital pathology, and multimodal models to aid in this differentiation and provides a table of reported diagnostic performance in published studies. AI can help improve risk classification and decrease unnecessary surgical procedures. However, existing evidence is insufficient, and its application should complement expert multidisciplinary evaluations, necessitating thorough validation to prevent delays in cancer diagnosis.
Citation: Salman A, Elewa A, Marwan A. Artificial intelligence in autoimmune pancreatitis: Differentiating immunoglobulin G4-related pancreatitis from pancreatic ductal adenocarcinoma. Artif Intell Gastroenterol 2026; 7(3): 123890
Autoimmune pancreatitis (AIP) is a rare yet significant fibroinflammatory condition of the pancreas, characterized by either diffuse pancreatic enlargement or a focal pancreatic mass. Type-1 AIP represents the pancreatic manifestation of immunoglobulin G4-related disease (IgG4-RD), while type-2 AIP is primarily pancreas-specific and is generally not associated with immunoglobulin G4 (IgG4)-positive plasma-cell infiltration or systemic IgG4-RD[1-3]. According to the International Consensus Diagnostic Criteria, AIP is divided into five major domains: Pancreatic imaging, serology, extrapancreatic organ involvement, histology, and response to steroid therapy[1]. This multimodal approach recognizes that none of the clinical, biochemical, radiological, or histological features is adequately reliable when used in isolation. Similarly, the inclusive criteria for IgG4-RD emphasize the relevance of clinical and radiological organ involvement, high serum IgG4 Levels, and characteristic pathological features, including extensive lymphoplasmacytic infiltration, storiform fibrosis, and obliterative phlebitis[2,3].
The main clinical problem is that AIP, especially localized or mass-forming AIP, might mimic pancreatic ductal adenocarcinoma (PDAC). Patients with either disease may present with obstructive jaundice, pancreatic enlargement or focal mass, pancreatic duct narrowing, weight loss, abdominal pain, and aberrant tumor markers. AIP is a steroid-responsive inflammatory process while PDAC is an aggressive cancer that requires rapid oncologic and surgical evaluation. The consequences of diagnostic mistake may be important: Overdiagnosis of cancer may lead to unnecessary pancreatectomy, and misclassification of PDAC as AIP may cause delay of potentially curative therapy. The current diagnostic pathways involve a judicious combination of pancreatic-protocol computed tomography (CT), magnetic resonance imaging (MRI) and magnetic resonance cholangiopancreatography (MRCP), endoscopic ultrasound (EUS), tissue acquisition, serum IgG4, carbohydrate antigen 19-9 (CA19-9), assessment for other organ involvement, and multidisciplinary clinical judgment[1,4,5].
Despite advances in cross-sectional imaging and EUS-guided tissue capture, distinguishing AIP from PDAC remains difficult in clinical practice. Elevated serum IgG4 can help with the diagnosis of type-1 AIP but is not diagnostic. Normal IgG4 levels can be seen in some patients with AIP, and high levels can also be seen in pancreatic cancer or other hepatopancreatobiliary disorders[2,3]. Recognizing classic features might provide diagnostic clues using histology, although pancreatic biopsy specimens may be limited by sampling error, insufficient tissue architecture, or difficulty demonstrating the full spectrum of IgG4-RD disease. Steroid responsiveness provides valuable insights; however, an empirical steroid trial should be approached with caution and implemented only after a thorough exclusion of malignancy. Inappropriate steroid exposure may obscure or postpone the diagnosis of PDAC[1,3]. These limitations establish a diagnostic context where supplementary decision-support tools could prove clinically beneficial.
Artificial intelligence (AI), especially machine learning, deep learning, and radiomics, has become a promising method for addressing complex image-based and multimodal diagnostic challenges in gastroenterology and pancreatic disease. Radiomics may convert routine medical images into high-dimensional quantitative data that reflects the lesion shape, intensity, texture, and geographic heterogeneity and has the ability to identify patterns that may not be readily apparent on visual assessment alone. In the context of AIP vs PDAC, CT radiomics and machine learning models have proven potential in discriminating these entities, especially in the differentiation of focal-type AIP from pancreatic cancer[6-8]. Other investigations have focused on MRI, diffusion-weighted imaging, positron emission tomography (PET)/CT radiomics and AI for the classification of pancreatic lesions, suggesting that computational image processing could complement conventional radiological interpretation[9-11].
However, the role of AI in this context should be approached with caution rather than mere enthusiasm. AIP is uncommon, and most existing AI studies are retrospective, single-center, and rely on small or imbalanced datasets[6-11]. Differences in scanners, acquisition protocols, segmentation methods, feature-selection pipelines, and reference standards may restrict reproducibility[6,9,10]. Furthermore, a model that demonstrates efficacy in a curated research cohort may not guarantee safety in patients presenting with indeterminate pancreatic masses, atypical imaging, incomplete serology, prior biliary stenting, or equivocal tissue sampling. Consequently, AI should not be considered as a substitute for radiologists, endosonographers, pathologists, or pancreatic multidisciplinary teams. Its function is mainly as a decision support tool to improve risk stratification, detection of subtle imaging signs, and more homogeneous integration of clinical, biochemical, radiological, endoscopic, and pathological data[6,9].
This review explores the changing importance of AI in distinguishing AIP, especially IgG4-related pancreatitis, from PDAC. The paper provides a critical review of imaging-based AI, radiomics, EUS-related applications, digital pathology, and multimodal diagnostic models, discussing limits, safety concerns, and requirements for clinical translation. To be transparent, we detail the strategy that was used for the literature search; when available, we provide an overview of the diagnostic performance of published models, and we distinguish explicitly between studies designed particularly for AIP and studies that deal with pancreatic lesions more generally. The principal applications of AI considered in this review are summarized in Table 1. The central argument posits that AI has the potential to resolve a significant diagnostic challenge in pancreatology, contingent upon its development and validation within clinically realistic frameworks that emphasize cancer safety, explainability, and multidisciplinary oversight.
Table 1 Potential applications of artificial intelligence in differentiating autoimmune pancreatitis from pancreatic ductal adenocarcinoma.
This is a narrative minireview. We searched PubMed, EMBASE, MEDLINE, Web of Science, and the Cochrane Library, supplemented by Google Scholar, from database inception to June 2026, restricting the final selection to English-language reports. Search terms combined controlled vocabulary and free text for “autoimmune pancreatitis”, “IgG4-related disease”, “pancreatic ductal adenocarcinoma”, “mass-forming pancreatitis”, “artificial intelligence”, “machine learning”, “deep learning”, “radiomics”, “convolutional neural network”, “endoscopic ultrasound”, and “digital pathology”. Reference lists of retrieved articles and relevant reviews were screened manually. We prioritized the following areas: (1) Studies utilizing AI, radiomics, or machine learning to differentiate AIP or mass-forming pancreatitis from PDAC; (2) Studies on AI relevant to the characterization of pancreatic lesions pertinent to this distinction; and (3) References on methodology and reporting standards for medical AI. Full-text access was restricted to open-access or retrievable articles, leading to the prioritization of such sources, while abstracts were utilized solely to verify bibliographic and quantitative information. Due to the limited number, variability, and primarily retrospective nature of the eligible studies, a formal meta-analysis was not conducted; instead, the reported performance is summarized descriptively (Table 2).
Table 2 Reported diagnostic performance of published artificial intelligence studies for differentiating autoimmune pancreatitis or mass-forming pancreatitis from pancreatic ductal adenocarcinoma.
The primary clinical challenge in AIP is accurately diagnosing the condition while simultaneously ensuring the exclusion of PDAC. This is especially challenging in focal or mass-forming AIP, where inflammatory pancreatic enlargement can create a distinct lesion, lead to ductal alterations, cause biliary obstruction, and present constitutional symptoms that mimic PDAC. Focal AIP can be seen as a pancreatic mass with obstructive jaundice, weight loss, abdominal discomfort, and cholestatic liver enzyme abnormalities, unlike diffuse AIP, which is usually characterized by diffuse enlargement and a capsule-like rim. The overlapping features elucidate why AIP continues to be a notable benign mimic of pancreatic malignancy[1,4,5,12].
The implications of diagnostic error are considerable. Misdiagnosis of AIP for PDAC may lead to unnecessary pancreatic resections, and misdiagnosis of PDAC for AIP may lead to delay in curative cancer therapy. Consequently, the International Consensus Diagnostic Criteria incorporate pancreatic imaging, serum IgG4 Levels, involvement of other organs, histological findings, and response to steroids, rather than depending on a singular characteristic[1]. Many patients do not meet all diagnostic criteria, and uncertainty persists when serology, imaging, or tissue sampling is incomplete or inconsistent[13].
Serology is a useful tool, yet it has limitations. Elevated serum IgG4 is indicative of type-1 AIP, especially when there is involvement of other organs. However, IgG4 Levels may be normal in AIP and can also be elevated in cases of pancreaticobiliary malignancy. Elevated CA19-9 levels may indicate a suspicion of PDAC; however, they can also be present in benign biliary obstruction, cholangitis, chronic pancreatitis, and, on occasion, AIP. Consequently, the interpretation of IgG4 and CA19-9 should occur in conjunction with imaging, histological findings, and clinical context[2-4,14].
Cross-sectional imaging is an important tool. CT and MRI/MRCP may reveal findings suggestive of AIP such as delayed enhancement, long-segment duct stricture, little proximal duct dilatation, diffuse or segmental enlargement, a capsule-like ring, and extrapancreatic IgG4-related illness. PDAC classically manifests as an ill-defined hypoenhancing tumor with abrupt duct cut-off, upstream duct dilatation, vascular invasion, pancreatic atrophy or metastasis. However, these are not absolute distinctions and diagnostic confusion remains for atypical or overlapping instances[5,15].
EUS gives high-resolution assessment of the pancreas and facilitates tissue biopsy in ambiguous masses. Although EUS can demonstrate diffuse hypoechoic enlargement, focal lesions, ductal constriction, or lymph-node abnormalities, morphology alone may not reliably distinguish focal AIP from PDAC. Histology may assist in the diagnosis of AIP by revealing lymphoplasmacytic sclerosing pancreatitis, IgG4-positive plasma cells, storiform fibrosis, or obliterative phlebitis and excluding malignancy. However, there may be limitations with EUS-guided sampling due to tiny fragments, sampling inaccuracy, poor tissue architecture, or uneven disease distribution[1,4,16].
Steroid responsiveness is a useful diagnostic tool but has important safety implications. A rapid clinical, biochemical, and radiological response is suggestive of AIP, but steroid treatments should not be used indiscriminately in patients with undetected pancreatic tumors. Malignancy should be carefully examined with imaging, tumor markers, EUS-guided tissue acquisition if appropriate, and multidisciplinary assessment, since inappropriate steroid exposure may delay definitive therapy of PDAC[1,4,17].
These constraints illustrate the clinical rationale for AI-assisted decision assistance. Differentiation of AIP from PDAC requires the combination of CT/MRI enhancement patterns, ductal morphology, lesion texture, diffusion characteristics, EUS appearance, serology, histology, and involvement of other organs. AI-based approaches, especially radiomics and multimodal machine-learning models, can assist by extracting quantitative imaging features and integrating clinical and imaging variables into structured risk estimates. Their value will depend on the extent to which they mitigate unnecessary surgery for AIP while not exacerbating delays in PDAC diagnosis[6-11].
IMAGING-BASED AI AND RADIOMICS
Imaging is the most advanced domain for AI-assisted differentiation between AIP and PDAC. The assessment of conventional radiography relies on visual analysis of enhancement, ductal morphology, pancreatic shape, vascular involvement, upstream atrophy, and extrapancreatic illness; however, these features often overlap in cases with localized AIP and PDAC. Radiomics allows the discovery of patterns that are not visually evident by converting routine photos into quantitative parameters that characterize lesion intensity, shape, margin, texture, and heterogeneity, hence enhancing the analysis of such images[18].
The evidence can be divided into research explicitly targeting AIP and PDAC and those addressing the larger issue of inflammatory (mass-forming) pancreatitis or pancreatic lesions in general. This difference matters because the performance of a curated AIP vs PDAC group may not be generalizable to indeterminate masses in the real world. Research concentrating on AIP primarily involves CT-based radiomics or machine-learning analyses of focal-type AIP[6-9]. Fewer studies utilize dual-time 2-[18F]fluorodeoxyglucose (FDG) PET/CT[10,19] and diffusion-weighted imaging[11]. Reported discrimination is notably high yet based on limited single-center cohorts: For instance, area under the curve (AUC) values of 0.975 for a CT random-forest model[6], 0.97 for a matched CT radiomics score[8], 0.967 for dual-time PET/CT radiomics[10], and 0.90 for a deep convolutional neural network (CNN) utilizing portal-venous CT[20]. The diagnostic performance, cohort sizes, modalities, validation strategies, and main limitations of the key studies are summarized in Table 2. A recent systematic review and meta-analysis of radiomics-based machine learning for differentiating pancreatitis from PDAC reported high pooled accuracy yet noted substantial heterogeneity and a lack of externally validated models[21].
The wide application of pancreatic-protocol CT in suspected PDAC has been investigated extensively in CT-based radiomics studies because it can provide multiphasic enhancement information. Although studies on AIP-specific CT radiomics are few, related studies on mass-forming pancreatitis support the idea that quantitative CT features may be useful in separating inflammatory pancreatic masses from PDAC. It is important to note that mass-forming (chronic) pancreatitis is not equivalent to AIP; thus, evidence from these cohorts is only indirectly relevant. Ye et al[22] established a mixed clinical-CT radiomics model to distinguish PDAC from focal mass-forming pancreatitis, which outperformed clinical-imaging- or radiomics-only models (test-set AUC 0.906). This allows the combination of radiomic and clinical data in inflammatory pancreatic cancer mimics.
Much attention has been paid to the focused type AIP. Anai et al[23] evaluated the effect of CT texture-based machine-learning analysis on the radiologists’ performance in distinguishing focal-type AIP from PDAC. They reported a standalone AUC of 0.920 for support-vector-machine, and an improvement in pooled reader AUC from 0.827 to 0.911 with the help of AI. Ziegelmayer et al[20] in a similar investigation, applied deep CNN-assisted feature extraction for discrimination and found that deep-feature model (AUC 0.90) outperformed traditional radiomics model. These results suggest that AI may be more successful as an adjunct to improve or standardise expert interpretation in difficult instances, rather than an independent diagnostic tool[20,23].
Radiomics-clinical models are a step forward in the creation of clinically relevant decision support systems. These algorithms can provide individualized diagnosis probability by combining radiomic signals with clinical or traditional imaging data. Qu et al[24] evaluated a radiomics-clinical combined model to differentiate pancreatic cancer from mass-forming pancreatitis and showed that the addition of clinical and radiomics data improved diagnostic discrimination (validation AUC 0.880). This approach better mimics real-world interdisciplinary evaluation than imaging-only categorization, while the risk of overfitting remains in small retrospective cohorts[24].
Diffusion weighted imaging and MRI may be helpful in AI-assisted distinction. MRI offers better soft tissue contrast, MRCP evaluates pancreaticobiliary ductal anatomy, and diffusion-weighted imaging reflects tissue and extracellular space cellularity. Choi et al[25] have shown that features on contrast-enhanced MRI and diffusion-weighted imaging may distinguish mass-forming AIP from PDAC and provide an imaging basis for future application of AI and radiomics using MRI-based data. Moreover, a diffusion weighted images radiomics follow-up model achieved an AUC of 0.962 in differentiating focal-type AIP from pancreatic cancer[11]. However, the MRI-based AI is affected by field strength, b-values, sequence parameters, motion artifacts, and reproducibility of segmentation[11,25].
The radiomics of PET/CT has been studied due to the possible increase in FDG uptake in inflammatory and malignant diseases. Liu et al[10] and Zhang et al[19] and evaluated the performance of 18F-FDG PET/CT radiomics to differentiate PDAC from AIP, reporting AUCs of 0.93 and 0.967, respectively. These results indicate that metabolic radiomics can provide information beyond the traditional visual interpretation of PET scans. This might be useful in situations that are complex or unclear, although PET/CT is not typically required in all patients with suspected AIP or PDAC[10,19].
Current imaging-based AI evidence, while promising, exhibits significant limitations. Many studies are retrospective and single-center, involving relatively small cohorts of AIP. Several of these studies report only the AUC without providing sensitivity or specificity, or they lack an independent test set (Table 2). Model performance can be influenced by variations in scanners, acquisition protocols, reconstruction settings, contrast timing, segmentation techniques, feature selection, and patient demographics. A model trained on confirmed AIP and confirmed PDAC may perform less well in real-world patients presenting with incomplete serology, biliary stents, equivocal cytology, small lesions, pancreatitis-related inflammatory changes, or alternative diagnoses including chronic pancreatitis, lymphoma, neuroendocrine tumors, or metastasis. Imaging AI should be a complementary tool for interdisciplinary interpretation, and the clinical benefit should be considered in terms of the capacity to safely lower the number of needless procedures without delaying the identification of PDAC[6-11,19,20,22-24].
EUS, PATHOLOGY, AND MULTIMODAL AI
EUS plays a critical role in evaluating indeterminate pancreatic masses by providing high-resolution imaging and allowing tissue acquisition. When comparing AIP with PDAC, EUS allows for the visualization of focal hypoechoic lesions, diffuse pancreatic swelling, ductal modifications, vascular connections, and lymph node anomalies. EUS-guided fine-needle biopsy is also useful to exclude malignancy and help in the diagnosis of lymphoplasmacytic sclerosing pancreatitis. AI-assisted EUS for AIP is not yet widely confirmed. However, studies on solid pancreatic lesions indicate that deep-learning systems may improve pancreatic tumor identification and characterization, thereby reducing operator-dependent variability[26,27].
AI-assisted EUS has a high potential benefit since conventional EUS interpretation mostly depends on the skill of the endosonographer. Deep-learning models can extract from EUS frames or films visual cues that are difficult to recognize by the naked eye and help differentiate malignant from inflammatory pancreatic lesions. A systematic review and meta-analysis showed a high pooled sensitivity and specificity of deep learning-assisted EUS in the identification of pancreatic cancers. However, it is crucial to highlight that most of the included studies were retrospective and not specific for AIP[26]. A recent deep learning study using EUS images exhibited promising classification between localized AIP and PDAC, representing the first EUS-based AIP-focused proof of concept, albeit it was performed in a small single-center cohort[28]. In the setting of suspected AIP, AI-EUS should be considered an adjunctive tool that may prompt additional tissue procurement or a multidisciplinary assessment, rather than a replacement for histology.
If imaging and serology are non-diagnostic, a tissue diagnosis is necessary. Eventually, digital pathology may assist in detecting malignant glands, inflammatory infiltrates, fibrosis patterns and distribution of IgG4-positive plasma cells. In the setting of AIP, traditional histological characteristics such as lymphoplasmacytic infiltration, storiform fibrosis, obliterative phlebitis and increased IgG4-positive plasma cells may be patchy or difficult to assess in tiny biopsies. Specific AI pathology models for AIP are currently limited, but the broader growth of computational pathology for pancreatic diseases suggests a potential future for automated quantification of inflammatory and malignant features, particularly when combined with expert pathological review. The recent development of general-purpose and vision-language pathology foundation models, trained on extensive slide archives, may facilitate this by promoting label-efficient identification of the histological features that distinguish type 1 AIP from PDAC[29,30].
Multimodal AI may be more clinically useful than single-modality models given that AIP diagnosis is inherently integrative. A full model would include CT or MRI radiomics, EUS image characteristics, histological information, serum IgG4 levels, CA19-9, bilirubin, ductal shape, additional organ involvement, and treatment response. Recent investigations on solid pancreatic lesions have demonstrated that the use of EUS pictures and clinical data increases diagnostic accuracy and facilitates a better human-AI collaboration across endoscopists with different degrees of expertise[31]. In line with this, a machine-learning model combining ultrasound and clinical features has been directly applied to the problem of AIP vs PDAC, further underlining that the combination of imaging and serological and clinical variables mirrors the actual diagnostic reasoning[32]. This technique is relevant to AIP since the main diagnostic difficulty is not only the classification of images but also the secure integration of inadequate and sometimes conflicting clinical information.
In summary, AI, digital pathology and EUS-based multimodal modeling are major advances beyond cross-sectional radiomics. But their relevance in AIP will need validation on a disease-specific basis, adequate representation of focused AIP, external testing and exact specification of clinical objectives. The best models will help determine when cancer has been reasonably excluded, when a repeat EUS-guided biopsy is indicated, and when a closely managed steroid treatment is indicated. Here, the AI should be a structured decision support layer for expert multidisciplinary care in pancreatic diseases, and not an autonomous diagnostic system[26,27,31].
CLINICAL IMPLEMENTATION, LIMITATIONS, AND FUTURE DIRECTIONS
The successful clinical application of AI to discriminate AIP from PDAC requires more than a high diagnostic accuracy on retrospective data sets. Considering the clinical consequences of missing PDAC, it is crucial to examine AI models on cancer safety, external validity, interpretability, and influence on real diagnostic decisions. Models must be validated on clinically relevant populations such as focal AIP, small PDAC, chronic pancreatitis, ambiguous pancreatic tumors, biliary obstruction, prior stenting, and incomplete serological or histological data. Transparency in reporting is essential and imaging-AI studies should conform to recognized reporting frameworks such as CLAIM and TRIPOD+AI[33,34].
From a practical perspective, it is beneficial to delineate the position of AI within the diagnostic pathway (Figure 1). AI is of limited utility when the diagnosis is already established; its greatest value lies in cases of indeterminate pancreatic masses that remain ambiguous following initial CT, serum IgG4, and CA19-9 assessments. In these patients, an AI-derived, calibrated probability could impact three decisions: (1) The choice to proceed directly to EUS-guided tissue acquisition instead of empirical management; (2) The justification for a carefully monitored steroid trial, contingent upon reasonable exclusion of malignancy; and (3) The necessity for urgent surgical-oncological referral. The patients who are most likely to benefit include those with focal or mass-forming lesions, atypical or discordant serology, or contraindications to surgery, as avoiding an unnecessary pancreatoduodenectomy offers significant advantages. AI output should never replace the multidisciplinary team, but should be seen in the context of a multidisciplinary team. Furthermore, data consistent with AIP should not supersede red-flag indications suggestive of malignancy.
Figure 1 Proposed artificial intelligence-supported diagnostic pathway for differentiating autoimmune pancreatitis from pancreatic ductal adenocarcinoma.
A patient presenting with a pancreatic mass, pancreatic enlargement, or obstructive jaundice undergoes conventional diagnostic assessment (computed tomography/magnetic resonance imaging/magnetic resonance cholangiopancreatography, endoscopic ultrasound with or without tissue acquisition, serology, and evaluation for other-organ involvement) in parallel with artificial intelligence-assisted analysis (radiomics/imaging analysis, endoscopic ultrasound image analysis, digital pathology, and multimodal integration). Both streams inform integrated risk stratification within a multidisciplinary team, which categorizes the patient as likely autoimmune pancreatitis (consider a carefully selected, closely monitored steroid trial), indeterminate (repeat endoscopic ultrasound-guided biopsy, short-interval imaging, or further review), or likely pancreatic ductal adenocarcinoma (urgent surgical or oncological pathway). The overarching goal is to reduce unnecessary surgery and invasive testing in autoimmune pancreatitis without delaying the diagnosis of pancreatic ductal adenocarcinoma. AIP: Autoimmune pancreatitis; CA19-9: Carbohydrate antigen 19-9; CT: Computed tomography; MRI: Magnetic resonance imaging; EUS: Endoscopic ultrasound; IgG4: Immunoglobulin G4; PDAC: Pancreatic ductal adenocarcinoma; MRCP: Magnetic resonance cholangiopancreatography; AI: Artificial intelligence.
A key drawback of the present evidence is that AIP is rare, which limits the size of the data set and increases the danger of overfitting. Many investigations are retrospective and single-institution research that rely on manually segmented lesions using carefully selected imaging procedures. Such conditions may lead to an overestimation of the model performance compared to ordinary practice when differences in scanner type, slice thickness, contrast timing, MRI sequence parameters, PET acquisition, and picture quality are important. Radiomics relies on several technical processes, such as segmentation, preprocessing, feature extraction, feature selection, and model training. Therefore, standardization of image biomarkers and radiomics methodology is essential before clinical application[35,36].
Interpretability constitutes a significant requirement. Clinicians are hesitant to accept a “black box” forecast for a patient with possibly resectable pancreatic cancer unless the AI output can be clarified and linked with imaging, EUS, serology, pathology and the whole clinical context. Explainable AI methodologies, like feature importance, heatmaps, attention maps, and calibrated risk scores, help doctors understand why a model prefers AIP or PDAC. However, explainability techniques need to be validated since visually compelling heatmaps do not necessarily show clinically meaningful reasoning[33,34].
The next step of future research should be a prospective clinical evaluation and not a technical proof of concept. Standardized data, including CT, MRI/MRCP, and EUS images, pathology results, IgG4 levels, CA19-9, bilirubin, other organ involvement, treatment response, and final diagnosis, must be collected from multicenter registries. AI systems must be validated externally across institutions and contrasted with experienced radiologists, endosonographers, and multidisciplinary teams. Early phase clinical evaluation should be done according to the DECIDE-AI principles, and interventional studies for the evaluation of AI-guided care should follow the SPIRIT-AI and CONSORT-AI reporting standards[37-39]. This is a demanding standard: A recent scoping review identified that randomized trials of AI in clinical practice are limited and methodologically inconsistent, with no existing trial specifically addressing AIP vs PDAC differentiation[40]. Broader experience with the clinical deployment of AI in gastroenterology indicates that regulatory approval, workflow integration, performance drift monitoring, and ethical and medicolegal considerations are equally important as diagnostic accuracy[41].
The optimal future model should be multimodal and feature an integrated workflow. Instead of merely categorizing an image as AIP or PDAC, it would offer a calibrated probability estimate, emphasize significant imaging and clinical features, indicate areas of uncertainty, and recommend clinically relevant next steps, including repeat EUS-guided biopsy, evaluation for other-organ IgG4-related disease, short-interval imaging, a cautious steroid trial, or urgent surgical-oncological referral[31,37-39]. Ultimately, the clinical usefulness of AI is in reducing unnecessary pancreatic operations and invasive testing in AIP and preserving prompt detection and treatment of PDAC[1,17,31].
CONCLUSION
AI could help solve a major diagnostic problem in pancreatology: AIP vs PDAC. Subtle imaging and clinical patterns commonly missed in conventional assessments may be identified via radiomics, deep learning, EUS-based image analysis, digital pathology and multimodal models. These methods may be especially useful in focal or mass-forming AIP, where there is significant clinical and radiological overlap with PDAC. AI should be seen as a decision support tool rather than a replacement for professional clinical judgment. The existing evidence is limited by small retrospective cohorts, single-center designs, uneven imaging procedures, and lack of significant external validation. Future progress will require prospective multicenter research, uniform imaging and pathology datasets, explainable models, and therapeutically relevant objectives. The ultimate value of AI will be its capacity to reduce unnecessary pancreatic surgeries and invasive testing in AIP without delaying the detection and treatment of PDAC.
Shimosegawa T, Chari ST, Frulloni L, Kamisawa T, Kawa S, Mino-Kenudson M, Kim MH, Klöppel G, Lerch MM, Löhr M, Notohara K, Okazaki K, Schneider A, Zhang L; International Association of Pancreatology. International consensus diagnostic criteria for autoimmune pancreatitis: guidelines of the International Association of Pancreatology.Pancreas. 2011;40:352-358.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 1338][Cited by in RCA: 1117][Article Influence: 74.5][Reference Citation Analysis (16)]
Umehara H, Okazaki K, Kawa S, Takahashi H, Goto H, Matsui S, Ishizaka N, Akamizu T, Sato Y, Kawano M; Research Program for Intractable Disease by the Ministry of Health, Labor and Welfare (MHLW) Japan. The 2020 revised comprehensive diagnostic (RCD) criteria for IgG4-RD.Mod Rheumatol. 2021;31:529-533.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 510][Cited by in RCA: 451][Article Influence: 90.2][Reference Citation Analysis (0)]
Li J, Liu F, Fang X, Cao K, Meng Y, Zhang H, Yu J, Feng X, Li Q, Liu Y, Wang L, Jiang H, Shao C, Lu J, Bian Y. CT Radiomics Features in Differentiation of Focal-Type Autoimmune Pancreatitis from Pancreatic Ductal Adenocarcinoma: A Propensity Score Analysis.Acad Radiol. 2022;29:358-366.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 5][Cited by in RCA: 24][Article Influence: 6.0][Reference Citation Analysis (0)]
Nakamura H, Fukuda M, Matsuda A, Makino N, Kimura H, Ohtaki Y, Nawa Y, Oyama S, Suzuki Y, Kobayashi T, Ishizawa T, Kakizaki Y, Ueno Y. Differentiating localized autoimmune pancreatitis and pancreatic ductal adenocarcinoma using endoscopic ultrasound images with deep learning.DEN Open. 2024;4:e344.
[RCA] [PubMed] [DOI] [Full Text][Cited by in RCA: 5][Reference Citation Analysis (0)]
Lambin P, Leijenaar RTH, Deist TM, Peerlings J, de Jong EEC, van Timmeren J, Sanduleanu S, Larue RTHM, Even AJG, Jochems A, van Wijk Y, Woodruff H, van Soest J, Lustberg T, Roelofs E, van Elmpt W, Dekker A, Mottaghy FM, Wildberger JE, Walsh S. Radiomics: the bridge between medical imaging and personalized medicine.Nat Rev Clin Oncol. 2017;14:749-762.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 4684][Cited by in RCA: 4323][Article Influence: 480.3][Reference Citation Analysis (20)]
Vasey B, Nagendran M, Campbell B, Clifton DA, Collins GS, Denaxas S, Denniston AK, Faes L, Geerts B, Ibrahim M, Liu X, Mateen BA, Mathur P, McCradden MD, Morgan L, Ordish J, Rogers C, Saria S, Ting DSW, Watkinson P, Weber W, Wheatstone P, McCulloch P; DECIDE-AI expert group. Reporting guideline for the early stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI.BMJ. 2022;377:e070904.
[RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)][Cited by in Crossref: 281][Cited by in RCA: 310][Article Influence: 77.5][Reference Citation Analysis (0)]
Cruz Rivera S, Liu X, Chan AW, Denniston AK, Calvert MJ; SPIRIT-AI and CONSORT-AI Working Group; SPIRIT-AI and CONSORT-AI Steering Group; SPIRIT-AI and CONSORT-AI Consensus Group. Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI extension.Nat Med. 2020;26:1351-1363.
[RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)][Cited by in Crossref: 453][Cited by in RCA: 471][Article Influence: 78.5][Reference Citation Analysis (7)]
