Immediate post-microwave ablation computed tomography model may better predict short-term response of stage I non-small cell lung cancer

Abstract

BACKGROUND

Microwave ablation (MWA) is indicated for treating inoperable early-stage non-small cell lung cancer.

AIM

To evaluate its therapeutic effect by developing multimodal models that integrate computed tomography (CT) radiomics features and simple clinical characteristics.

METHODS

This study enrolled 50 patients with stage I non-small cell lung cancer who were treated with MWA. Corresponding CT images were collected before and after the MWA, and radiomics features were extracted using the 3D-slicer software platform. In the study, statistical analyses were performed using R studio software. Firstly, radiomics features related to local tumor progression (LTP) at 6 months were screened from the pre-MWA and immediate post-MWA CT radiomics features by the max-relevance and min-redundancy features. This study aimed to evaluate its therapeutic effect by developing multimodal models that integrate CT radiomics features and simple clinical characteristics. Based on the selected radiomics features, prediction models for 6-month LTP were constructed. These models’ performance was evaluated by the area under the receiver operating characteristic (ROC) curve (AUC).

RESULTS

Three pre-MWA CT radiomics features and three post-MWA CT radiomics features were confirmed to be correlated with the LTP. For the ROC curves that only covered the pre-MWA CT radiomics, AUC was 0.689. For the ROC curves integrating the pre-MWA CT radiomics plus simple clinical characteristics, AUC was 0.849. For the ROC curves that only covered the immediate post-MWA CT radiomics features, AUC was 0.849. For the ROC curves integrating the immediate post-MWA CT features plus simple clinical characteristics, AUC was 0.907.

CONCLUSION

This exploratory preliminary analysis revealed that the predictive model integrating immediate post-microwave-ablation CT radiomic features and clinical indicators yields superior prognostic performance relative to the model constructed solely from preoperative CT radiomic features.

Key Words: Microwave ablation; Non-small cell lung cancer; Stage I; Radiomics; Computed tomography; Treatment response; Prediction model; Short-term efficacy

Core Tip: This exploratory study constructed predictive models based on preoperative and immediate post-microwave ablation computed tomography radiomic features as well as simple clinical characteristics from patients with stage I non-small cell lung cancer (NSCLC), and verified that the immediate postoperative multimodal model may exhibit favorable predictive performance for the risk of local tumor progression at 6 months. This model can assist clinicians in rapid patient risk stratification, guide the formulation of individualized surveillance schedules, facilitate early salvage local interventions, and coordinate subsequent systemic therapy. It effectively improves the refined management paradigm after microwave ablation for early-stage NSCLC and provides novel imaging evidence to support precise implementation of minimally invasive comprehensive treatment for NSCLC.

INTRODUCTION

The advance of high-resolution computed tomography (CT) has contributed to the continuous improvement of the detection rate of early lung cancer in recent decades[1]. Despite the surgery being the preferred treatment for patients with early-stage cancer, it is not applicable for all patients. For example, lobectomy is the first-line therapy method for stage I non-small cell lung cancer (NSCLC)[2], but it may impair cardiopulmonary function in elderly patients or those with multiple comorbidities. Against this backdrop, local treatment options including thermal ablation and stereotactic body radiation therapy are considered viable options for inoperable patients[3,4]. The Cardiovascular and Interventional Radiological Society of Europe 2024 guidelines: Ablation is recommended as a treatment option for medically inoperable patients with stage IA (T1a/b N0 M0) peripheral NSCLC, provided that a minimum 3-mm ablation margin can be achieved around major bronchi and blood vessels.

In the study by Narsule et al[5], thermal ablation was capable of effectively controlling or treating the stage IA NSCLC for inoperable patients. Ni et al[6] found the median overall survival (OS) was 64.2 months. The OS rate was 99% at 1 year, 75.6% at 3 years, and 54.1% at 5 years[6]. According to previous studies, thermal ablation can feasibly and effectively serve for primary and secondary pulmonary malignant tumors, yielded 2-year OS rates comparable to those achieved by surgical resection and stereotactic body radiation therapy based on the propensity score matching[7,8]. For example, the RAPTURE study, a prospective multicenter clinical trial that treated 13 patients of early-stage NSCLC using radiofrequency ablation, reported a 2-year OS rate of 75%[9]. The trial ascertained the tolerable toxicities of radiofrequency ablation (RFA). Furthermore, a prospective multicenter phase II trial[10] published in 2018 came to a similar conclusion.

Lung tumor ablation has demonstrated proven efficacy, yet the development of ablation therapies for pulmonary tumors has progressed slowly. The short-term efficacy and OS outcomes of thermal ablation for early-stage lung cancer were required further enhancement. In order to improve the local control rate of ablation therapy of early-stage NSCLC patients as well as reduce their recurrence rate. Meanwhile, to enhance the promotion and adoption of lung tumor ablation therapy in clinical settings, it is in urgent need to predict their prognosis by corresponding models, and adopt remedial rational programs to improve the therapeutic effect.

CT-based radiomics is capable of extracting multiple quantitative imaging features from CT scans, and these images assist in capturing the tumor heterogeneity, and serve for the prediction and assessment of treatment response as potential imaging biomarkers. Specifically, CT-based radiomics has always been applied to effectively monitor NSCLC in terms of the tumor progression and the therapeutic efficacy[11-13]. Cutting-edge imaging technology benefits the quantitative analysis of the medical radiomic images, and can evaluate patients’ prognosis effect based on these images[14-16]. According to Wu et al[16], radiomics features regarding primary tumor or peritumoral region could offer some treatment reference for NSCLC in terms of the survival prediction, distant metastasis and treatment response. They developed a radiomic method to examine the tumorous and peritumoral areas in CT images, meanwhile designing a non-invasive biomarker to achieve the effective differentiation of responders to therapy and the accurate classification of the survival outcomes.

Pre-ablation and post-ablation radiomic features exhibited the difference of tumorous and peritumoral areas in evaluating therapeutic outcomes of lung tumor ablation. While CT-based radiomics has become pivotal in clinical decision-making, few studies have leveraged pre-ablation or immediate post-ablation CT imaging combined with simple clinical characteristics to predict therapeutic outcomes of ablation for NSCLC. This study aims to integrate pre-operative CT and/or post-ablation CT imaging plus simple clinical characteristics to predict treatment efficacy in early-stage NSCLC, and evaluate the performance of each model.

MATERIALS AND METHODS

Patients

The study obtained the approval of the Institutional Review Board of the Ethics Committee of Central Hospital Affiliated to Shandong First Medical University. We obtained the written informed consent of all participants with lung tumors before they received CT-guided percutaneous microwave ablation (MWA).

Inclusion criteria: (1) Adult patients (aged ≥ 18 years) with lung tumors who were admitted to our department to receive MWA; (2) With malignant lung tumors of NSCLC diagnosed in histological or cytological levels; (3) Re-examined by chest enhanced CT 4-6 weeks following MWA, with results as the evaluation baselines; (4) With maximum tumor diameter < 5 cm; and (5) With organs of lung functioning normally.

Exclusion criteria: (1) Patients who received prior radiotherapy or other local therapies before or after MWA; (2) Missing/incomplete follow-up data; and (3) Absence of pathological confirmation. In total, the study included 50 NSCLC patients who had received MWA from November 2019 to December 2023 based on the retrospective analysis. The study flow chart is shown in Figure 1.

Figure 1 The study flow chart. NSCLC: Non-small-cell lung cancer; MWA: Microwave ablation; CT: Computed tomography.

Imaging protocol and MWA

Before ablation, routine non-contrast chest CT scans were performed for all patients using a Philips Brilliance Big Bore large-bore CT scanner with helical scanning mode. The scan parameters were set as follows: Tube voltage 120 kV, X-ray tube current 350 mA, slice thickness 5 mm, rotation time 0.75 seconds, pitch 1, single collimation width 1.5 mm, total coverage length 24 mm. Images were reconstructed with convolution kernel B and the third-generation iterative reconstruction algorithm iDose³. A radiologist excelling in thoracic intervention adopted a set protocol for such surgery, confirming the puncture site and the most proper entrance path of the microwave antenna after observing the CT images. The tumor location and size were taken into account to adjust the power settings and the ablation times as per the producer’s protocol (4-7 minutes active tip at 55W or 50W). Lidocaine was injected into the puncture site for inducing the local pleural anesthesia. The radiologist performed MWA using a medical microwave applicator (Yigao MWA system, China) Normal saline was perfused via a peristaltic pump at 60 mL/minute to protect normal tissue along the proximal antenna shaft. For small lesions (less than 3.0 cm) or medium tumor (3.0-5.0 cm), single or dual needle applicators were used to achieve complete ablation. The completion of MWA was followed by the removal of the ablation antenna, and CT scans were repeated immediately to assess ablation-associated complications and the complete elimination of the tumor.

Assessment of radiomics changes

Flowchart of radiomic analysis is shown in Figure 2. Delineation method was selected considering whether it was available and useful in clinical practices. One experienced radiologist singly took charge of manually contouring all lesions, delineated the region of interest (ROI) on lung window settings (window width 1000 HU, window level -600 HU). The validation of ROIs was performed by a senior radiologist. The segmented 3D volume was subjected to comparative analysis. We extracted the textural features from the ROIs after analyzing the pre-MWA and post-MWA non-contrast-enhanced CT images (a 5-mm slice thickness). We obtained sufficient information from larger tumors. Body posture did not exert a big influence in the CT scanning process. Then, these CT images were imported to process the ROI structure on the 3D-slicer software platform (version 4.8.1). A total of 847 radiomic features were automatically extracted, such as the gray-level co-occurrence matrix, Gray Level Run Length Matrix, Intensity Direct, Neighbor Intensity Difference and Shape, and their calculation also relied on 3D-slicer software.

Figure 2 Flowchart of radiomic analysis. A-C: Representative axial slices of pre-microwave ablation computed tomography scans, with yellow contours delineating the primary lung tumor; D-F: Representative axial slices of immediate post-microwave ablation computed tomography scans, yellow contours mark the ablation zone after thermal ablation. MWA: Microwave ablation; CT: Computed tomography.

Evaluation of the local efficacy

The study took the short-term evaluation of the 6-month LTP as the outcome of interest, which referred to situation that a new contrast-enhanced lesion appeared in the follow-up visit after a successful surgery was documented. Efficacy was assessed according to the Response Evaluation Criteria in Solid Tumors version 1.1 (RECIST 1.1), categorizing results as complete response, partial response, stable disease, and progressive disease. Two radiologists with diverse experience levels independently took charge of evaluating the post enhanced CT images after the ablation, while blinded to patients’ clinical data and follow-up outcomes. Classification of complications was conducted taking into account the occurrence time and the imaging manifestations, and relevant grading followed the Society of Interventional Radiology (SIR) International Working Group on Image-Guided Tumor Ablation. Chest enhanced CT examination was conducted 4-6 weeks following the surgery, with the results being the disease evaluation baseline.

Statistical analysis

We used max-relevance and min-redundancy (mRMR) method for feature selection, retaining the three most relevant CT radiomics, which essentially achieved feature dimensionality reduction. R software mRMRe package was used to conduct the analyses. All feature data were expressed as mean ± SD. Logistic regression was used to build the prediction models, taking the 6-month LTP as the dependent variable (progression = 1, no progression = 0). The predictors included the radiomics features selected by the mRMR method and simple clinical characteristics (age, gender, tumor type, and stage).

AUC analysis was conducted for assessing the discriminating power of models. AUC stands for area under the receiver operating characteristic (ROC) curve. It is a single-number summary of the ROC curve, and it is one of the most important evaluation metrics for checking the performance of a classification model, especially for binary classification problems. Other performance measures include sensitivity, specificity, accuracy, positive-predictive value, negative-predictive value. These measures were estimated using the Youden statistics. All the analyses were conducted in R studio software (version 4.3.3).

RESULTS

Patient characteristics

The study enrolled 50 patients in total, with 24 males and 26 females aged from 35 to 92 (average age: 68.56 years ± 11.68 years old). Table 1 lists their demographic information. Pathological types related to these patients included adenocarcinoma (78%) and squamous cell carcinoma (22%). All patients had stage I NSCLC, with 47 patients in T1 stage and 3 patients in T2 stage. 10% of patients (n = 5) developed local tumor progression in 6 months after MWA.

Table 1 Patient demographic and clinical characteristics.

Clinical features	n (%)
Gender
Female	26 (52)
Male	24 (48)
Age, median (range) (years) 69 (50-84)
≤ 65	15 (30)
65-80	28 (56)
≥ 81	7 (14)
Age (mean ± SD)	68.56 ± 11.68
Pathological type
Adenocarcinoma	39 (78)
Squamous cell carcinoma	11 (22)
T stage
T1	47 (94)
T2	3 (6)
Smoker
Yes	21 (42)
No	29 (58)

Correlation between radiomics features analysis and LTP

The study adopted the mRMR feature selection method. Relevant analysis relied on the R software mRMRe package. Logistic regression was used to construct the ROC curves of the prognostic model.

Three features of pre-MWA CT radiomics (wavelet-high-high-low/glszm/large area low gray-level emphasis, wavelet-low-low-high/gldm/dependence entropy and wavelet-high-low-low/glszm/large area high gray-level emphasis) were confirmed to present a significant correlation with the 6-month LTP, with their mean and SD values in different prognosis groups listed in Table 2. The ROC curves with regard to the 6-month LTP model integrating the pre-MWA CT radiomic features, the predictive ability was evaluated according to the ROC curves, with the AUC of 0.689 (Figure 3A). The ROC curves with regard to the 6-month LTP model integrating the pre-MWA CT radiomic features and simple clinical characteristics, the predictive ability was evaluated according to the ROC curves, with the AUC of 0.849 (Figure 3B).

Figure 3 Receiver operating characteristic curves for predicting 6-month local tumor progression using four different prediction models. A: The receiver operating characteristic (ROC) curves of prediction models of 6-month local tumor progression (LTP) based on pre-microwave ablation (MWA) computed tomography (CT) features. The ROC curves were used to evaluate the models, and the areas under the curve area under the ROC curve (AUC) was 0.689; B: The ROC curves of prediction models of 6-month LTP based on pre-MWA CT features and simple clinical characteristics. The ROC curves were used to evaluate the models, and the areas under the curve AUC was 0.849; C: The ROC curves of prediction models of 6-month LTP based on post-MWA CT features. The ROC curves were used to evaluate the models, and the areas under the curve AUC was 0.849; D: The ROC curves of prediction models of 6-month LTP based on post-MWA CT features and clinical variables. The ROC curves were used to evaluate the models, and the areas under the curve AUC was 0.907. MWA: Microwave ablation; CT: Computed tomography; AUC: Areas under the curve area under the receiver operating characteristic curve.

Table 2 Results of feature selection for pre-microwave ablation computed tomography radiomics predicting 6-month local tumor progression.

Image features	No progress		Progress
	Mean	SD	Mean	SD
Wavelet-HHL/glszm/LALGLE	14320.75	30365.52	108574.87	235623.06
Wavelet-LLH/gldm/DE	6.14	0.44	6.46	0.33
Wavelet-HLL/glszm/LAHGLE	1698565.88	2731697.10	9608440.19	12803613.18

LALGLE: Large area low gray level emphasis; DE: Dependence entropy; LAHGLE: Large Area high gray level emphasis; HHL: High-high-low.

Three features of post-MWA CT radiomics (wavelet-high-high-high/glcm/inverse variance, original/shape/maximum 2D diameter slice and wavelet-low-low-low/glcm/maximum probability) were confirmed to present a significant correlation with the 6-month LTP, with their mean and SD values in different prognosis groups listed in Table 3. The ROC curves with regard to the 6-month LTP model integrating the post-MWA CT radiomic features, the predictive ability was evaluated according to the ROC curves, with the AUC of 0.849 (Figure 3C). The ROC curves with regard to the 6-month LTP model integrating the post-MWA CT radiomic features and simple clinical characteristics, the predictive ability was evaluated according to the ROC curves, with the AUC of 0.907 (Figure 3D).

Table 3 Results of feature selection for post-microwave ablation computed tomography radiomics predicting 6-month local tumor progression.

Image features	No progress		Progress
	Mean	SD	Mean	SD
Wavelet-HHH/glcm/IV	0.48	0.01	0.49	0.01
Original/shape/M2DDC	43.05	17.87	67.60	39.65
Wavelet-LLL/glcm/MP	0.004	0.003	0.013	0.009

HHH: High-high-high; IV: Inverse variance; M2DDC: Maximum 2D diameter column; LLL: Low-low-low; MP: Maximum probability.

Comparison between predictive models

Comparative model performance evaluated in Table 4. Among these models, the post-CT plus simple clinical characteristics model demonstrates the best discriminative ability with the highest AUC of 0.907. The pre-MWA CT plus simple clinical characteristics model achieves the highest overall Accuracy of 0.90, along with the highest specificity of 0.91. In terms of sensitivity, all models except the pre-MWA CT radiomics model (0.60) reach a value of 1.00, indicating their ability to identify all true positive cases. The negative-predictive value of the last two groups of models all reaches 1.00, making their negative predictions highly reliable. Overall, models that combine radiomics and simple clinical characteristics generally exhibit better performance, with the post-CT radiomics plus simple clinical characteristics model standing out as the best in comprehensive performance.

Table 4 Comparative performance of five predictive models.

Performance measures	Models
	Pre-MWA CT radiomics	Pre-MWA CT + clinical variables	Post-MWA CT radiomics	Post-MWA CT + clinical variables
Youden	0.47	0.71	0.71	0.73
Sensitivity	0.60	0.80	1.00	1.00
Specificity	0.87	0.91	0.71	0.73
Accuracy	0.84	0.90	0.74	0.76
NPV	0.95	0.98	1.00	1.00
PPV	0.33	0.50	0.28	0.29
True positive	3	4	5	5
False positive	6	4	13	12
True negative	39	41	32	33
False negative	2	1	0	0
AUC	0.689	0.849	0.849	0.907

MWA: Microwave ablation; CT: Computed tomography; NPV: Negative predictive value; PPV: Positive-predictive value; AUC: Area under the receiver operating characteristic curve.

DISCUSSION

Lung cancer related screening and techniques have developed greatly, promoting the application of MWA in the treatment of early-stage NSCLC patients, of which the complete ablation rate reaches 80%-90%[17]. For patients with stage I NSCLC who are elderly, complicated with underlying comorbidities, or contraindicated for surgical resection, MWA has emerged as one of the preferred curative local therapeutic modalities[6,18]. Existing clinical studies have validated that MWA achieves favorable safety and efficacy profiles for inoperable stage I NSCLC, with a 5-year OS rate of up to 54.1%[6]. Most treatment-related adverse events are mild, enabling effective local tumor control and durable long-term survival benefits for patients. Prospective clinical data further corroborate the clinical utility of MWA specifically for inoperable stage I NSCLC patients aged ≥ 70 years. In this elderly cohort, the 1-year post-procedure OS rate reaches 99.0% and the 1-year progression-free survival rate is 93.7%, with zero treatment-related mortality. MWA therefore serves as a safe and minimally invasive therapeutic option for vulnerable elderly patients[18]. Chest enhanced CT examination, conducted 4-6 weeks following the surgery, aims at assessing the local treatment efficacy[19]. However, short-term postoperative imaging assessment alone cannot predict the long-term risk of local tumor progression, which limits precise guidance for individualized post-ablation management. Therefore, we constructed prediction models of pre-MWA or immediate post-MWA CT radiomics integrated simple clinical characteristics to predict the therapeutic effect of 6 months LTP and guide timely salvage therapy after tumor ablation, thereby increasing the complete ablation rate. It bears great clinical significance for early identification of patients at high risk of progression after MWA and improvement of the comprehensive treatment system for stage I NSCLC, and provides objective imaging evidence to formulate individualized surveillance and intervention strategies. We found that immediate post-MWA CT radiomics plus simple clinical characteristics model exhibited favorable predictive performance.

Accumulating evidence has demonstrated that radiomics enables in-depth mining of numerous quantitative microscopic textural, density, and morphological features imperceptible to the naked eye on medical images. Such features can accurately reflect the microscopic pathological status of tumors and ablation zones, and radiomics has been widely applied in efficacy evaluation, prognostic prediction, and risk stratification for lung cancer[20-23]. Chemoradiotherapy, targeted therapy and immunotherapy can alter tumor heterogeneity[24], and such alterations can be quantitatively reflected via radiomic features, which serve as independent reference indicators for predicting therapeutic response[25]. At present, studies seldom pay attention to the prediction of lung cancer patients’ response to ablation therapy by using radiomics features. Compared with conventional imaging evaluation, radiomics techniques exhibit superior performance in predicting ablation therapeutic response. MWA completely eradicates local tumor tissue through high-temperature coagulative necrosis, triggering characteristic remodeling of intralesional density and texture architecture. Such dynamic postoperative imaging changes are strongly correlated with tumor residue and long-term risk of local progression. Empirical findings indicate that MWA resulted in the variation in the intratumoral density changes, and the corresponding prognostic value in predicting treatment response and LTP was confirmed. Sun et al[20] found that CT radiomics stratified survival outcomes in metastatic NSCLC after MWA, outperforming conventional prognostic markers. This suggests that pre- and post-ablation CT radiomics provide distinct prognostic information[26].

The present study revealed that the predictive model solely constructed based on preoperative CT radiomic features exhibited limited performance for predicting local tumor progression at 6 months post-ablation, with an AUC of only 0.689, which fails to meet the demands of clinical risk assessment. We screened three preoperative radiomic features (wavelet-high-high-low/glszm/large area low gray-level emphasis, wavelet-low-low-high/gldm/dependence entropy and wavelet-high-low-low/glszm/large area high gray-level emphasis) significantly associated with post-procedural tumor progression, mainly reflecting the original tumor grayscale, texture and regional distribution signatures. Although these parameters could partially reflect baseline tumor heterogeneity, they were incapable of capturing pathological remodeling lesions induced by MWA, which accounted for the insufficient predictive power of the preoperative-only model. In contrast, the multimodal model integrating immediate postoperative CT radiomic features and simple clinical characteristics achieved markedly improved predictive efficacy, with an AUC of 0.907, demonstrating excellent capacity for early risk stratification. The core postoperative radiomic features screened in this study included gray-level co-occurrence signatures, morphological parameters and texture entropy of ablation lesions. These metrics precisely captured a series of specific imaging alterations after MWA, such as tumor necrosis, inflammatory exudation and tissue repair. They could objectively quantify the completeness of ablation and the underlying risk of tumor recurrence and progression, serving as reliable quantitative biomarkers for forecasting short-term local progression.

This area of completed ablation of lung cancer was subjected to zonal tissue necrosis from the central ablation cavity to the peripheral parenchyma, effusion and congestion in histological level[26], and the CT manifestation was ground-glass opacity (GGO)[27], with an obvious change on the imaging features. In 2023, Huang et al[28] predicted immediate response of RFA in treating colorectal cancer lung metastases, extracting the pre-RFA and post-RFA radiomics features from CT scans. Complete ablation status was significantly correlated with nine pre- and immediate post-ablation CT radiomic features (pre-RFA and immediate post-RFA CT). On this basis, our analysis identified three post-MWA CT radiomics features (wavelet-high-high-high/glcm/inverse variance, original/shape/maximum 2D diameter slice, and wavelet-low-low-low/glcm/maximum probability) that significantly correlated with 6-month LTP. After comprehensive validation, we found the 6-month LTP prognostic model of the post-CT radiomics plus simple clinical characteristics stood out the best comprehensive performance. The AUC of 6-month LTP prediction model was 0.907.

The 6-month local tumor progression prediction model established in this study delivers distinct clinical guiding value within the full-cycle comprehensive therapeutic framework for NSCLC, effectively compensating for the deficiencies of existing post-ablation management protocols. Previous studies have confirmed that patients with inoperable stage I NSCLC still carry a certain risk of local progression following MWA. Notably, patients with stage IB disease demonstrate significantly poorer prognosis than those with stage IA, constituting a high-risk cohort for postoperative recurrence and progression. Conventional post-ablation surveillance protocols adopt uniform reexamination intervals and assessment criteria, leading to insufficient intervention for high-risk individuals and excessive follow-up for low-risk patients, accompanied by inefficient allocation of medical resources.

Relying on the predictive model developed herein, risk stratification can be performed immediately after MWA. For patients at high risk of progression, an intensified surveillance schedule can be implemented with shortened intervals of chest CT scans to facilitate early detection and diagnosis of incipient tumor recurrence. Meanwhile, individualized early intervention strategies can be formulated: Salvage local therapies including repeat ablation and local radiotherapy can be promptly administered for lesions with suspected residue or high progression risk, maximizing the local tumor control rate. Conversely, simplified surveillance schedules with reduced reexamination frequency can be adopted for low-risk patients to alleviate their medical burden.

Furthermore, the risk prediction capacity of our model can be integrated into the systemic full-course treatment system for NSCLC, refining the comprehensive diagnosis and treatment paradigm combining local ablation and systemic therapy. For patients with advanced NSCLC who develop local progression after MWA and are ineligible for further curative local interventions, anlotinib, a novel multi-target tyrosine kinase inhibitor, exerts anti-tumor effects by suppressing tumor angiogenesis and proliferative signaling pathways, serving as a standard later-line therapeutic regimen for advanced NSCLC. Multiple clinical trials[29,30] have validated that third-line and beyond anlotinib therapy markedly prolongs progression-free survival and OS, elevates disease control rates, and presents manageable adverse reactions with favorable clinical applicability. Our model enables early identification of patients with elevated progression risk, providing predictive evidence for the timely initiation of subsequent systemic targeted therapies such as anlotinib. This achieves full-process optimization of care: Risk prediction after local ablation → early salvage local treatment → subsequent systemic therapy upon progression, further improving the comprehensive management system for minimally invasive treatment of early-stage NSCLC.

In addition, oriented toward practical clinical demands, this study abandons cumbersome multi-dimensional clinical features and constructs a convenient and efficient predictive model centered on postoperative CT radiomic features supplemented by simple clinical covariates, matching the clinical scenario of rapid bedside assessment. Compared with intricate machine learning algorithms, the logistic regression model adopted in this study features stable parameters and robust reproducibility, facilitating clinical translation and widespread implementation. It offers a novel quantitative tool for evaluating therapeutic response after MWA for stage I NSCLC in hospitals at all levels.

Several limitations of the present study should be acknowledged. First, this is a single-center retrospective exploratory study with a limited sample size and a small number of outcome events, which may predispose the model to overfitting and restrict the generalizability of statistical findings. The insufficiency of data compromised the model's credibility, prompting additional AUC testing for each model. After testing and analysis, model of the post-CT radiomics plus simple clinical characteristics stands out the best comprehensive performance. Second, due to sample size limitations, we were unable to employ AI-based approaches for predictive model development, and we lacked an independent validation cohort for rigorous performance verification. We employed logistic regression to develop our predictive model, as its simpler, parametric nature offers greater robustness and a significantly lower risk of overfitting compared to more advanced machine learning models like SVMs or Random Forests. Third, the screening of clinical variables in this study centered on clinical practicality. We prioritized basic, readily accessible clinical features and did not incorporate refined indicators such as ablation parameters, inflammatory biomarkers, and pulmonary function metrics. The variable screening framework can be optimized in follow-up research to further improve the predictive accuracy of the model. Fourth, only internal dataset analysis was performed in the current study without an independent external validation cohort, so the generalizability of the model remains to be further verified. In future work, we will expand the sample size, enroll multicenter cases, attempt to optimize model architecture with deep learning algorithms, and conduct external cohort validation to enhance model stability and clinical applicability. In our subsequent research, we collected multicenter cohort data using immediate post-ablation CT images from patients with primary lung cancer and pulmonary metastases for exploratory validation. The validation results confirmed the prominent advantages of postoperative CT images for predicting ablation efficacy. Our ultimate goal is to establish a streamlined predictive framework based on concise and efficient indicators to optimize efficacy assessment after ablation.

CONCLUSION

In conclusion, this exploratory study constructed predictive models based on preoperative and immediate post-MWA CT radiomic features as well as simple clinical characteristics from patients with stage I NSCLC, and verified that the immediate postoperative multimodal model may exhibit favorable predictive performance for the risk of local tumor progression at 6 months. This model can assist clinicians in rapid patient risk stratification, guide the formulation of individualized surveillance schedules, facilitate early salvage local interventions, and coordinate subsequent systemic therapy. It effectively improves the refined management paradigm after MWA for early-stage NSCLC and provides novel imaging evidence to support precise implementation of minimally invasive comprehensive treatment for NSCLC.

ACKNOWLEDGEMENTS

The authors would like to thank all the patients and authors involved in this study.