Deep learning model based on automated CT image segmentation for predicting the optimal radiotherapy protocol in limited-stage small cell lung cancer patients: a multicenter study

Introduction

According to global data, lung cancer remains the leading cause of cancer-related deaths. Small cell lung cancer (SCLC) accounts for approximately 15% of all lung tumors, with about one-third of SCLC patients diagnosed at the localized stage (1-4). Clinically, this localized presentation is referred to as limited-stage SCLC (LS-SCLC), which is broadly defined as disease confined to one hemithorax and its corresponding regional lymph nodes that can be safely encompassed within a single definitive radiotherapy field. Concurrent chemoradiotherapy has long been the standard treatment for LS-SCLC, with the optimal radiation dose and fractionation regimen for thoracic radiotherapy remaining a subject of ongoing attention and debate (5). Results from two meta-analyses indicate that adding radiotherapy to chemotherapy improves median survival, 3-year survival, and local control rates (6). INT 0096 research demonstrated that compared to once-daily radiotherapy [once daily (QD), 1.8 Gy per fraction], twice-daily radiotherapy [twice daily (BID), 1.5 Gy per fraction] significantly improved 5-year overall survival (OS) (26% vs. 16%) (7). Moreover, previous studies have demonstrated that SCLC cancer cells exhibit high sensitivity to radiotherapy, with tumor cells undergoing exponential death even at low doses (8). Consequently, twice-daily radiotherapy combined with chemotherapy has become the standard treatment for LS-SCLC (9).

Conventional fractionation regimens are more convenient to implement in clinical practice, with patients requiring only one treatment session per day, thereby improving patient compliance (10). Furthermore, the toxicity of radiotherapy (such as radiation pneumonitis and esophagitis) may be lower compared to hyperfractionated regimens (11). A substudy of the CALGB 30610 trial suggested that conventional fractionation radiotherapy may offer superior quality of life outcomes compared to hyperfractionated radiotherapy, particularly regarding fatigue and dysphagia (12). However, some studies have also found that the local control rate of conventional fractionation may be inferior to that of hyperfractionation, particularly at the standard dose of 45 Gy (13). The hyperfractionated radiotherapy regimen (45 Gy/30 fractions) is the standard treatment for LS-SCLC, with its efficacy confirmed by multiple studies demonstrating significant superiority over conventional radiotherapy regimens (14). The BID regimen’s shorter overall treatment duration may reduce tumor cell repopulation and improve local control rates (15). Moreover, some studies have explored the efficacy of high-dose hyperfractionated regimens (such as 60 Gy/40 fractions or 54 Gy/30 fractions). Initial results indicate that these approaches are safe and may offer potential survival benefits (16). However, the hyperfractionated regimen carries a higher incidence of acute toxicity, which may impact patients’ quality of life and treatment completion rates. It also places greater demands on medical resources and requires more flexible scheduling for patients.

Therefore, it remains unclear which patients may benefit from which radiotherapy regimen. This study proposes a two-stage automated prediction framework for radiotherapy regimens, integrating “image segmentation-response prediction”. It demonstrates that combining “automated segmentation and point-prompt segmentation” with a ResNet18 prediction model—constructed using computed tomography (CT) images and voxel-level radiomics feature maps—effectively predicts optimal radiotherapy regimens for LS-SCLC patients undergoing chemotherapy and radiotherapy. We present this article in accordance with the TRIPOD reporting checklist (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-1-0023/rc).

Methods

Patients selection

A retrospective analysis was conducted on data from LS-SCLC patients who received radiotherapy combined with chemotherapy at two research centers between January 2018 and December 2024. Center A (Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences) serves as the training cohort, while Center B (Xiaogan Central Hospital) serves as the external validation cohort. All eligible patients had pathologically confirmed SCLC; all received either hyperfractionated radiotherapy (BID) or conventionally fractionated radiotherapy (QD); all patients received 4–6 cycles of standard chemotherapy. Patients who received antitumor treatments other than radiotherapy or chemotherapy, experienced treatment interruptions, had concurrent tumors, had incomplete information, or were lost to follow-up were excluded from the study. This complete-case analysis strategy was chosen to maintain the integrity of the deep learning (DL) feature extraction process, ensuring that the model learned from a fully curated dataset without the need for data imputation. Figure 1 shows the study design flowchart. Detailed treatment protocols are provided in Appendix 1. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Ethics Committees of Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences (No. SDTHEC202509011) and Xiaogan Central Hospital (No. XGSZXYYLL-EC-20251210). Patient consent was waived as the retrospective nature of this study.

Figure 1 Overview of the study design and analytical workflow. (A) This study incorporated data from two distinct centers. (B) VISTA3D architecture. It comprises two branches sharing the same image encoder. If the user provides a class prompt within the 127 supported classes, the top auto branch activates ready-made automatic segmentation. If the user provides a 3D point-click prompt, the bottom interactive branch activates interactive segmentation. If both branches are activated, the fusion module uses the interactive result to edit the automatic result. (C) The DL model employs ResNet18. (D) Four widely recognized quantitative metrics in medical image segmentation were selected: Dice similarity coefficient, 95% Hausdorff distance, mean surface distance, and volume similarity. These metrics were used to calculate the consistency between the three segmentation results and manual segmentation. The primary evaluation metric for the DL model is the Harrell C-index. The study also employs NPV, PPV, ACC, SEN, SPE, calibration curves, and decision curves to assess model predictive performance. (E) To deepen understanding of the model’s decision basis and enhance its clinical credibility, this study employed Grad-CAM. This heatmap visually reveals the contribution of different spatial regions within the input image to the final risk score. (F) Finally, a decision pathway based on this study was established. 3D, three-dimensional; ACC, accuracy; C-index, concordance index; CT, computed tomography; DL, deep learning; Grad-CAM, gradient-weighted class activation mapping; LS-SCLC, limited-stage small cell lung cancer; NPV, negative predictive value; PPV, positive predictive value; SEN, sensitivity; SPE, specificity.

CT image acquisition and preprocessing

Prior to radiotherapy, CT imaging was performed on all enrolled patients. The resulting images were subsequently transmitted to the Picture Archiving and Communication System (PACS) and retrieved in the standard DICOM format. In instances where multiple imaging sessions were conducted for a given patient, the specific scan temporally closest to the commencement of treatment was selected for analysis. To ensure uniformity across the dataset, a series of standardized preprocessing steps was implemented. Recognizing that variations in voxel spacing can introduce confounding factors, spatial normalization was applied. Specifically, this was achieved through a fixed-resolution resampling technique, whereby all images were uniformly resampled to an isotropic voxel dimension of 1 mm³. Following spatial alignment, intensity normalization was performed using z-score standardization (also referred to as zero-mean normalization) to standardize the range of pixel values, with lung window parameters defined by a width of 1,500 Hounsfield units (HU) and a level of −600 HU. To address potential discrepancies in target definition arising from inter-institutional practices, manual segmentation of the primary gross tumor volume was conducted. This critical task was undertaken collaboratively by two senior radiation oncology radiologists, each possessing a decade of clinical experience, utilizing ITK-SNAP software (version 4.4.0; available at https://www.itksnap.org/). These meticulously crafted manual delineations were subsequently adopted as the reference standard, against which the performance of the VISTA3D-based automated segmentation and subsequent feature extraction pipeline was rigorously evaluated.

VISTA3D segmentation DL model

This investigation utilized the VISTA3D framework, a unified architecture for volumetric medical image segmentation built upon a SegResNet encoder and a sliding-window inference mechanism (17) (Figure 2A). By incorporating three-dimensional (3D) super-voxel distillation knowledge alongside a four-stage training regimen, the framework is capable of automated delineation across 127 anatomical classes, facilitates interactive 3D refinement, and supports zero-shot segmentation tasks. The core reliability assessment of this study is centered on a comparative analysis of three distinct segmentation methodologies supported by VISTA3D, with their outputs evaluated against manually annotated reference standards using an external test cohort. The specific methodologies under evaluation are delineated as follows: (I) fully automated segmentation, which utilizes the shared SegResNet backbone and an auto-branching mechanism to execute end-to-end delineation of target structures without the need for manual input. (II) Point-prompt-based, interactive segmentation. (III) A hybrid approach integrating fully automated segmentation followed by point-prompt refinement. To quantify the congruence between the outcomes of these three methods and the manual gold standard, four widely acknowledged quantitative metrics in the field of medical image analysis were selected: the Dice similarity coefficient, the 95^th percentile Hausdorff distance, the mean surface distance, and the volume similarity index. A comprehensive exposition of the protocols governing this evaluation is available in Appendix 2.

Figure 2 Automatic segmentation and habitat analysis. (A) VISTA3D architecture. It comprises two branches sharing the same image encoder. If the user provides a class prompt within the 127 supported classes, the top automatic branch activates pre-trained automatic segmentation. If the user provides a 3D point prompt, the bottom interactive branch activates interactive segmentation. If both branches are activated, the fusion module uses the interactive results to edit the automatic results. (B) The VISTA3D-based automated segmentation architecture not only segments the tumor lesion itself but also constructs constructed a 3 mm peritumoral ring, dividing both the tumor and peritumoral ring into multiple subregions. The K-means clustering algorithm is employed to partition all voxels within the tumor into K distinct clusters. (C) Comparison of segmentation performance among three segmentation models. (D) Radar chart of DL model prediction performance. (E) Calibration curve for evaluating the predictive performance of DL models. (F) Decision curve for evaluating the predictive performance of DL models. (G) Violin plot of DL score distribution for BID and QD patients. 3D, three-dimensional; ACC, accuracy; AUC, area under the curve; BID, twice daily; DL, deep learning; PPV, positive predictive value; QD, once daily; SEN, sensitivity; SPE, specificity.

Habitat segmentation

A 3-mm peritumoral ring was delineated around the primary tumor, and both the tumor and its surrounding ring were further partitioned into multiple subregions. To account for intra-tumoral heterogeneity, we applied unsupervised clustering to classify voxels within the tumor. Initially, principal component analysis was utilized to reduce the dimensionality of the feature space, which helped eliminate redundant information and improve computational efficiency. Then, K-means clustering was adopted to assign tumor voxels into K distinct clusters, with each cluster representing a distinct radiological habitat. The optimal number of clusters was identified through a combination of the elbow method and silhouette coefficient. The resulting peritumoral expansion and its subdivisions are depicted in Figure 2B.

3D ResNet18 dual-branch efficacy prediction model

This study employs ResNet18 as the feature extraction backbone. Addressing the clinical need for selecting radiotherapy regimens for LS-SCLC, an end-to-end DL prediction model was constructed. The core objective of the model is to simultaneously assess the progression-free survival (PFS) risk for the same patient following two distinct radiotherapy regimens (conventional fractionation vs. hyperfractionation) based on pre-treatment 3D CT images, thereby identifying the optimal personalized treatment plan for specific patients. To achieve this goal, the model employs a unique dual-branch architecture. This architecture shares a common foundation in a 3D feature extraction network, which is responsible for learning general deep features from CT images that correlate with tumor biological behavior and treatment response. Subsequently, these shared features are fed in parallel into two prediction branches that share identical architecture but have independent parameters. Each branch specializes in predicting treatment outcomes under a specific radiotherapy regimen, ultimately generating a continuous risk score. The advantage of this design is that it allows the model to directly compare the expected benefit differences between two treatment strategies for a specific patient based on the same feature representation. This avoids the potential feature inconsistency issues that could arise from training two independent models separately, making direct comparisons between treatment regimens more reliable. To balance the retention of pre-trained knowledge with the adaptability to CT image features, we employed a layered freezing strategy. In the final configuration, shallow layers of the model were frozen, with only two deep layers participating in gradient updates. We implemented the aggregator module based on AttentionMIL, which comprises two layers of multi-layer perceptrons (MLPs) and a Softmax activation function.

To optimize an end-to-end DL model for predicting the efficacy of radiotherapy regimens in LS-SCLC patients, this study designed a composite loss function—L_total that integrates three core objectives: precise regimen-specific survival prediction, enhanced ability to distinguish between regimen efficacies, and temporal plausibility of the predicted survival function.

The total loss function consists of three core components: L_total = λ₁L_survival + λ₂L_contrast + λ₃L_consistency.

Where λ₁=1.0, λ₁=0.3, λ₃=0.1. The treatment-specific survival prediction loss L_survival is based on the negative log-likelihood of the Cox proportional hazards model, optimized separately for conventional fractionated radiotherapy and hyperfractionated radiotherapy, aiming to ensure that the model-predicted risk scores accurately reflect the PFS risk ranking of patients. To enhance the model’s ability to clearly distinguish which treatment regimen is superior for specific patients, a ternary marginal comparison loss function (L_contrast) was introduced. This approach constructs patient triplets and constrains the differences in their treatment advantage scores, thereby strengthening the model’s discriminative power. Additionally, the time-consistency loss (L_consistency) imposes monotonic non-increasing and smoothness constraints on the predicted survival function, ensuring it aligns with clinically observed survival time dynamics. This composite loss function is minimized using the AdamW optimizer, coupled with a cosine annealing learning rate scheduling strategy. The final model outputs a treatment preference score (ASI(X)=1|T|∑t∈T[S^B(t|X)−S^A(t|X)]) to guide clinical decision-making. Detailed loss function design specifications are provided in Appendix 3.

The primary evaluation metric for DL models is the area under the curve (AUC). The study also employs negative predictive value (NPV), positive predictive value (PPV), accuracy, sensitivity, specificity, calibration curves, and decision curves to assess model predictive performance. Detailed protocols are provided in Appendix 4.

To deeply understand the basis for model decisions and enhance their clinical credibility, this study adopted a combined approach of gradient-weighted class activation mapping (GW-CAM) and feature importance analysis. For each patient’s CT image input, we computed and generated a protocol-specific GW-CAM heatmap. These regions may contain information potentially associated with tumor radiobiological behavior, serving as imaging surrogates for phenotypes such as proliferation activity and hypoxia levels. However, these interpretations remain speculative in the absence of direct molecular validation. These regions may contain information related to tumor radiobiological behavior, such as proliferation activity and hypoxia levels. Overlaying the heatmaps generated for conventional fractionated radiotherapy and hyperfractionated radiotherapy allows clinicians to visually compare the model’s focus areas when evaluating these two distinct treatment regimens. This enables them to understand why the model might recommend one approach over the other.

Statistical analysis

For statistical analysis, the Python statsmodels (version 3.7.12) module was utilized, and a P value <0.05 was deemed statistically significant. We analyzed the differences between groups using the Student’s t-test or Mann-Whitney U test for continuous variables and the Chi-squared test or Fisher’s exact test for categorical variables. Kaplan-Meier analysis was used to evaluate OS and PFS. Censoring was defined as the absence of a recorded event (progression or death) at the date of the last follow-up or the end of the study period. All censored cases were treated as non-informative, and the censoring distribution was accounted for in the survival estimates.

Results

Patient characteristics

The study ultimately included 446 patients, with 314 in the training cohort and 132 in the external validation cohort. By the end of follow-up, 305 (68.39%) patients experienced disease progression, and 180 (40.36%) patients experienced death. Among the two cohorts, 156 (49.68%) and 70 (53.03%) patients were aged ≥60 years, respectively; 227 (72.29%) and 96 (72.73%) patients were male, respectively. The majority of patients in both groups were stage T2, comprising 106 (33.76%) and 55 (41.67%) patients, respectively. A total of 171 (54.46%) and 65 (49.24%) patients in the two groups received prophylactic cranial irradiation, respectively (Table 1).

Performance of the 3D automated segmentation

The external test queue of Center B was employed to evaluate the segmentation performance of three core segmentation methods. The comparison results demonstrate that the point-prompt-enhanced automatic segmentation method effectively resolves misclassification issues inherent in standalone automatic segmentation while maintaining segmentation efficiency. The Dice similarity coefficient for this method is 0.86±0.17, representing a 14% and 5% improvement over independent automatic segmentation and point-prompt segmentation, respectively. It also achieves significant enhancements in 95% Hausdorff distance, average surface distance, and volume similarity metrics. Detailed results are shown in Figure 2C. Furthermore, this method significantly reduces the manual interaction cost compared to independent point-prompt segmentation, making it the chosen implementation for subsequent automated segmentation-based optimal radiotherapy regimen prediction tasks.

Model performance

The AUC for PFS prediction by the DL model was 0.88 [95% confidence interval (CI): 0.71–0.96], with an accuracy of 0.84 (95% CI: 0.71–0.93), a NPV of 0.88 (95% CI: 0.77–0.93), PPV was 0.74 (95% CI: 0.64–0.88), sensitivity was 0.77 (95% CI: 0.66–0.89), and specificity was 0.84 (95% CI: 0.74–0.94) (Figure 2D). The calibration curve of the DL model showed a clear trend (Figure 2E). The model’s discriminatory ability and clinical applicability were evaluated through decision curve analysis (Figure 2F). This study generated treatment preference scores for all patients using the DL model. Based on the empirical analysis of the training data, patients with scores >0.5 were recommended for BID radiotherapy, while those with scores ≤0.5 were recommended for QD radiotherapy. There were 221 and 225 patients in these two groups, respectively, with a significant difference in scores between the two groups (Figure 2G, Tables S1,S2).

This study also conducted interpretability analysis on the DL model, validating its decision logic from multiple dimensions: in the slice heatmaps of samples across the X, Y, and Z orthogonal planes, high-attention regions (with normalized weights approaching 1) precisely anchored tumor cores or pathologically specific tissue areas, while background tissues predominantly exhibited low weights (approaching 0). Moreover, the high-weight regions across samples and dimensions highly overlap with clinically relevant observation areas, confirming that model decisions rely on task-related core features. This provides direct visual support for the interpretability of prediction results (Figure 3). Among patients with a score >0.5, patient A (score =0.7125) receiving BID radiotherapy demonstrated significantly longer PFS compared to patient B (score =0.9255) receiving QD radiotherapy [16.1 (95% CI: 12.5–16.8) vs. 8.6 (95% CI: 5.3–13.5)]. Among patients with a score ≤0.5, there was no significant difference in PFS between patient C (score =0.0919) receiving BID radiotherapy and patient D (score =0.3121) receiving QD radiotherapy [11.4 (95% CI: 8.6-14.2) vs. 12.8 (95% CI: 8.7–16.4)]. However, the incidence of grades 3–4 acute esophagitis was significantly lower in patients receiving QD radiotherapy (16.0% vs. 9.4%, P<0.01).

Figure 3 Model interpretability analysis. Among patients with a score >0.5, patient A (score =0.7125) receiving BID radiotherapy demonstrated significantly superior PFS compared to patient B (score =0.9255) receiving QD radiotherapy [16.1 (95% CI: 12.5–16.8) vs. 8.6 (95% CI: 5.3–13.5)]. Among patients with a score ≤0.5, there was no significant difference in PFS between patient C (score =0.0919) receiving BID radiotherapy and patient D (score =0.3121) receiving QD radiotherapy [11.4 (95% CI: 8.6–14.2) vs. 12.8 (95% CI: 8.7–16.4)]. The attention weight heatmap visually represents the model’s focus on lesions. BID, twice daily; CI, confidence interval; HR, hazard ratio; PFS, progression-free survival; QD, once daily.

Among patients with a score >0.5, PFS analysis demonstrated that patients receiving BID radiotherapy showed significantly superior outcomes compared to those receiving QD radiotherapy [hazard ratio (HR) =0.55; 95% CI: 0.40–0.76; P<0.001] (Figure 4A). OS analysis revealed that patients receiving BID radiotherapy demonstrated significantly superior outcomes compared to those receiving QD radiotherapy (HR =0.39; 95% CI: 0.26–0.59; P<0.001) (Figure 4B). In the forest plot of subgroup analysis, the results from each subgroup align with the primary analysis, demonstrating a significant prognostic benefit for patients receiving BID radiotherapy (Figure 4C,4D). In patients with a score ≤0.5, there was no significant difference in PFS between those receiving BID radiotherapy and those receiving QD radiotherapy (HR =1.12; 95% CI: 0.87–1.45; P=0.83) (Figure 4E). Similarly, no significant difference was observed in the OS analysis (HR =1.14; 95% CI: 0.83–1.41; P=0.85) (Figure 4F). Given the wider CIs in this subgroup, these results suggest a lack of demonstrated survival benefit for hyperfractionation in this specific imaging phenotype, supporting the potential for treatment de-intensification to reduce toxicity. The clinical decision pathway established based on this study is detailed in Figure 5.

Figure 4 Kaplan-Meier survival analysis for patients identified by DL models as suitable for BID and QD radiotherapy groups. (A) Kaplan-Meier survival analysis of PFS for patients identified by the model as suitable for BID radiotherapy. (B) Kaplan-Meier survival analysis of OS for patients identified by the model as suitable for BID radiotherapy. (C) Subgroup analysis of PFS. (D) Subgroup analysis of OS. (E) Kaplan-Meier survival analysis of PFS for patients identified by the model as suitable for QD radiotherapy. (F) Kaplan-Meier survival analysis of OS for patients identified by the model as suitable for QD radiotherapy. BID, twice daily; CCRT, concurrent chemoradiotherapy; CI, confidence interval; DL, deep learning; EC, etoposide and carboplatin; ECOG, Eastern Cooperative Oncology Group; EP, etoposide and cisplatin; HR, hazard ratio; OS, overall survival; PFS, progression-free survival; QD, once daily.

Figure 5 Clinical decision pathway. The DL model based on ResNet18 established in this study classifies patients as suitable for hyperfractionated radiotherapy or conventional fractionated radiotherapy. For patients identified by the model with a score >0.5, the study considers that this group exhibits higher aggressiveness and is therefore more suitable for hyperfractionated radiotherapy regimens. For patients identified by the model as having a score ≤0.5, the study considers that this group exhibits relatively lower aggressiveness, with no significant difference between the two treatment regimens. However, conventional radiotherapy demonstrates a lower incidence of acute adverse events. BID, twice daily; DL, deep learning; QD, once daily.

Discussion

Radiotherapy combined with chemotherapy is the standard treatment regimen for LS-SCLC patients. However, the optimal radiotherapy regimen for patients remains unclear. Identifying which patients may benefit from specific radiotherapy regimens is crucial, not only to improve patient prognosis but also to reduce the incidence of adverse events. Therefore, this study first achieved automated tumor segmentation for LS-SCLC patients by combining automatic segmentation with point-prompt segmentation technology, establishing a DL model for predicting treatment efficacy. The study findings demonstrate that the model can identify patients who may significantly benefit from hyperfractionated radiotherapy. Conversely, for patients predicted by the model to derive no benefit from hyperfractionation, conventional fractionation protocols exhibit superior toxicity control. This provides valuable insights for precision radiotherapy decision-making in LS-SCLC.

In the treatment of LS-SCLC, although concurrent chemoradiotherapy is the standard regimen, the optimal fractionation pattern remains controversial (5,13). INT 0096 study established the 45 Gy/30 fractions hyperfractionated regimen as one of the standard treatment options, exhibiting survival advantages over the conventional fractionation regimen (7). However, the higher acute toxicity associated with hyperfractionated regimens (such as esophagitis and pneumonia), along with their greater demands on medical resources and patient compliance, urgently necessitate the exploration of more optimized individualized treatment options (11,12,18).

The proposed “automatic segmentation-treatment efficacy prediction” two-stage framework achieves an end-to-end analysis from raw CT images to personalized treatment recommendations. First, utilizing the advanced VISTA3D segmentation model, we achieved high-precision and highly efficient automatic segmentation of tumor regions, providing a reliable foundation for subsequent analysis. While the proposed framework supports a fully automatic end-to-end workflow from raw CT images, the integration of point-prompt technology essentially provides a high-precision semi-automatic refinement option. This allows clinicians to maintain ultimate control over the segmentation quality, balancing operational efficiency with the necessity for expert-level accuracy in personalized treatment planning. Second, the constructed dual-branch 3D ResNet18 model represents the core innovation of this research. Through its unique design combining shared feature extraction and treatment-specific branches, this model can directly learn and compare the expected efficacy of two radiotherapy regimens from pre-treatment CT images, outputting a continuous risk score (19). The model demonstrated strong predictive performance in an external validation cohort (AUC =0.88).

In the subgroup of patients recommended for hyperfractionated radiotherapy by the DL model (ASI >0.5), those who actually received hyperfractionated treatment demonstrated significantly superior PFS and OS compared to those receiving conventional fractionation (HR were 0.55 and 0.39, respectively). This finding is consistent with previous studies supporting the superiority of hyperfractionated radiotherapy in specific patient populations (7,14,20). In the subgroup of patients identified by the model as having a low predicted benefit from hyperfractionation (ASI ≤0.5), there was no statistically significant difference in PFS or OS between patients who actually received BID vs. QD radiotherapy. This indicates that within this ASI-defined subgroup, the choice of fractionation regimen does not materially affect survival outcomes. Separately, as a post-hoc clinical observation, we found that among patients in this low-ASI subgroup, those who received QD radiotherapy had a significantly lower incidence of grades 3–4 acute esophagitis compared to those who received BID (9.4% vs. 16.0%, P<0.01). It is critical to note that this toxicity finding was not learned or predicted by the model—the model was trained exclusively on PFS data and does not incorporate any toxicity endpoints. Rather, this observation emerged from clinical follow-up after patients were stratified by the model. When considered alongside the model’s PFS-based recommendation, this toxicity data can provide additional clinical context to support shared decision-making, particularly for patients where efficacy is equivalent between regimens. This finding is consistent with the results of the CALGB 30610 study (12) and aligns with the recent emphasis on balancing efficacy with patient-reported outcomes in lung cancer treatment (21). Therefore, this model not only identifies potential beneficiaries of hyperfractionated radiotherapy but also selects patient groups who may be better suited for conventional fractionation regimens.

A critical distinction in personalized oncology is whether a model is prognostic or predictive. While prognostic models estimate the overall outcome of a disease, predictive models—such as the one developed in this study—identify patients likely to have a differential response to a specific treatment. Our data support the predictive utility of the ASI score, as evidenced by the significant survival benefit of BID radiotherapy exclusively in the high-score group (ASI >0.5, HR =0.55), while the low-score group (ASI ≤0.5) showed no survival difference but benefited from the lower toxicity of QD radiotherapy. This indicates that the DL framework decodes imaging phenotypes specifically related to radiation sensitivity and repopulation dynamics.

A notable characteristic of our DL model is its heavy reliance on imaging-derived features. While this supports our goal of establishing a fully automated workflow, we acknowledge that established clinical covariates—such as Eastern Cooperative Oncology Group (ECOG) performance status, smoking history, and chemotherapy timing—carry significant prognostic weight. Our internal benchmarking suggests that the DL model (ASI score) captures biological information (e.g., intratumoral heterogeneity and aggressiveness) that is not fully represented by standard clinical variables like T-stage. Specifically, the ASI score identifies a “radiological phenotype” that predicts benefit from hyperfractionation, whereas clinical variables often function more as prognostic markers of overall fitness rather than predictive markers for specific fractionation benefits.

The interpretability analysis enhances the model’s clinical credibility. GW-CAM reveals that the model primarily focuses on the tumor core region during decision-making, aligning with the clinical priorities of radiation oncologists. This indicates that the model does not operate as a “black box”, but rather makes inferences based on imaging features potentially associated with tumor biological behavior and radiotherapy response. Intratumoral spatial heterogeneity has been demonstrated to correlate with treatment resistance and prognosis across multiple cancers (22,23). The DL model based on ResNet18 established in this study classifies patients to the group suitable for BID radiotherapy and the group suitable for QD radiotherapy. CT images of patients A and B reveal tumors exhibiting significant intra-tumoral heterogeneity, characterized by irregular margins, spiculation, and non-uniform density. The GW-CAM heatmap indicates that the model’s attention is primarily focused on specific high-density subregions within the tumor core and at the tumor invasion front. These high-attention areas may correspond to regions of high cellular density, active proliferation, or hypoxic microenvironments—biological features known to be associated with radiation resistance and requiring intensified radiotherapy regimens. The model’s precise localization of these subregions supports its recommendation for BID radiotherapy, as this regimen biologically overcomes accelerated repopulation and hypoxia-induced radioresistance in aggressive SCLC more effectively. In contrast, CT images of patients C and D showed that although the tumors were of SCLC type, the entire tumor mass exhibited relatively more uniform density. The GW-CAM heatmap revealed diffuse low-intensity distribution of attention weights across the entire tumor region, without significant focus on any specific subregion. This pattern indicates the model failed to identify discrete high-risk subregions warranting treatment intensification. Biologically, this may suggest that despite histological SCLC characteristics, the tumor lacks specific intratumoral microenvironments (e.g., hyperproliferative clones, hypoxic cores) that confer sensitivity to segmentation scheme variations.

The current study has several notable limitations. First, the retrospective design of this study is a primary limitation, introducing unavoidable treatment selection bias. Because the choice between hyperfractionated and conventionally fractionated radiotherapy was clinician-driven, it may correlate with unmeasured clinical confounders, such as patient performance status, baseline pulmonary reserve, or specific tumor bulk characteristics. While our model targets imaging-based biological features, these clinical factors could influence both treatment assignment and observed outcomes. Therefore, prospective randomized validation is essential to decouple these confounders and confirm the independent predictive power of the ASI score. Second, while the overall cohort size of 446 patients is reasonable for LS-SCLC, the sample size within specific subgroups—particularly after stratification by both the ASI score and the radiotherapy fractionation regimen—is relatively limited. Although the high rate of progression events (68.39%) provides a degree of statistical robustness, these subgroup analyses may still be underpowered to detect subtle differences. Consequently, the reported HRs and their corresponding 95% CIs, especially those demonstrating non-significance in the low-score group (ASI ≤0.5), should be interpreted with caution. These findings should be regarded as hypothesis-generating. Third, certain important clinical variables, including ECOG status, nodal burden, and smoking history, were not explicitly incorporated into the current DL architecture. The absence of these variables may limit the model’s ability to account for the patient’s overall physiological reserve. Future research should focus on developing multi-modal fusion models that synergize deep imaging features with comprehensive clinical datasets to provide a more holistic decision-support system. Finally, the biological underpinnings of the DL-derived features and habitats have not yet been directly validated, which warrants future investigation integrating radiogenomics or spatially-resolved pathology to decode the model’s decision logic.

Conclusions

The DL model developed in this study, based on automatic segmentation combined with point-prompt technology, provides a tool for predicting PFS benefit from BID vs. QD radiotherapy in LS-SCLC patients. The model stratifies patients into those likely to derive significant PFS benefit from BID and those for whom PFS is equivalent between regimens. In the latter group, post-hoc clinical observation of reduced esophagitis with QD provides additional context to support shared decision-making. This two-stage framework—integrating automated segmentation with response prediction—demonstrates the feasibility of using imaging-based DL to inform personalized treatment discussions, though prospective validation is needed before clinical implementation.

Acknowledgments

We thank each patient for permitting the use of their clinicopathological data for this study.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-1-0023/rc

Data Sharing Statement: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-1-0023/dss

Peer Review File: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-1-0023/prf

Funding: This work was supported by the Young Elite Sponsorship Program of Shandong Provincial Medical Association (No. 2024-GJ-0059), the Research Project of Shandong Provincial Medical Association (No. YXH2024YS053), the Special Research Fund for Standardized Management of Bone Marrow Inhibition Discipline Construction of China Anti-Cancer Association (No. CETSDCORP2772-01), and the Wu Jieping Foundation (No. 320.6750.2025-06-192).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-1-0023/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Ethics Committees of Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences (No. SDTHEC202509011) and Xiaogan Central Hospital (No. XGSZXYYLL-EC-20251210). Patient consent was waived as the retrospective nature of this study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Sung H, Ferlay J, Siegel RL, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin 2021;71:209-49. [Crossref] [PubMed]
2. Han J, Hua Y, Wu X, et al. Chemoradioimmunotherapy versus chemoimmunotherapy or chemoradiotherapy in extensive-stage small-cell lung cancer: A retrospective analysis of survival outcomes. Precis Radiat Oncol 2025;9:295-303. [Crossref] [PubMed]
3. Zhao H, Qi Y, Zhang L, et al. Thoracic radiotherapy timing and prognostic factors in elderly patients with limited-stage small cell lung cancer. Precis Radiat Oncol 2024;8:14-21. [Crossref] [PubMed]
4. Rudin CM, Ismaila N, Hann CL, et al. Treatment of Small-Cell Lung Cancer: American Society of Clinical Oncology Endorsement of the American College of Chest Physicians Guideline. J Clin Oncol 2015;33:4106-11. [Crossref] [PubMed]
5. Faivre-Finn C, Snee M, Ashcroft L, et al. Concurrent once-daily versus twice-daily chemoradiotherapy in patients with limited-stage small-cell lung cancer (CONVERT): an open-label, phase 3, randomised, superiority trial. Lancet Oncol 2017;18:1116-25. [Crossref] [PubMed]
6. Warde P, Payne D. Does thoracic irradiation improve survival and local control in limited-stage small-cell carcinoma of the lung? A meta-analysis. J Clin Oncol 1992;10:890-5.
7. Turrisi AT 3rd, Kim K, Blum R, et al. Twice-daily compared with once-daily thoracic radiotherapy in limited small-cell lung cancer treated concurrently with cisplatin and etoposide. N Engl J Med 1999;340:265-71. [Crossref] [PubMed]
8. Jeremic B, Shibamoto Y, Acimovic L, et al. Initial versus delayed accelerated hyperfractionated radiation therapy and concurrent chemotherapy in limited small-cell lung cancer: a randomized study. J Clin Oncol 1997;15:893-900. [Crossref] [PubMed]
9. Dingemans AC, Früh M, Ardizzoni A, et al. Small-cell lung cancer: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up Ann Oncol 2021;32:839-53. [Crossref] [PubMed]
10. Luo J, Song J, Xiao L, et al. Simultaneous integrated dose reduction intensity-modulated radiotherapy effectively reduces cardiac toxicity in limited-stage small cell lung cancer. Cancer Biol Med 2023;20:452-64. [Crossref] [PubMed]
11. Zhan TY, Zhang T, Wang WQ, et al. Simultaneous Integrated Dose Reduction Radiation Therapy Versus Conventional Fractionated Radiation Therapy With Chemotherapy for Limited-Stage Small Cell Lung Cancer in the Modern Treatment Era: A Phase 3, Randomized Clinical Trial (SDR-RT trial). Int J Radiat Oncol Biol Phys 2026;124:692-701. [Crossref] [PubMed]
12. Ganti AK, Fruth B, Rimner A, et al. Quality of life outcomes in patients participating in the CALGB 30610 trial (CALGB 70702): Alliance. Cancer 2025;131:e35663. [Crossref] [PubMed]
13. Deek MP, Haigentz M, Jabbour SK. Waiting for Big Changes in Limited-Stage Small-Cell Lung Cancer: For Now, More of the Same. J Clin Oncol 2023;41:2326-30. [Crossref] [PubMed]
14. Yu J, Jiang L, Zhao L, et al. High-dose hyperfractionated simultaneous integrated boost radiotherapy versus standard-dose radiotherapy for limited-stage small-cell lung cancer in China: a multicentre, open-label, randomised, phase 3 trial. Lancet Respir Med 2024;12:799-809. [Crossref] [PubMed]
15. Grønberg BH, Killingberg KT, Fløtten Ø, et al. High-Dose Versus Standard-Dose Twice-Daily Thoracic Radiotherapy in Limited-Stage SCLC: Final Survival Data, Long-Term Toxicity, and Relapse Patterns in a Randomized, Open-Label, Phase II Trial. J Thorac Oncol 2025;20:1108-19. [Crossref] [PubMed]
16. Grønberg BH, Killingberg KT, Fløtten Ø, et al. High-dose versus standard-dose twice-daily thoracic radiotherapy for patients with limited stage small-cell lung cancer: an open-label, randomised, phase 2 trial. Lancet Oncol 2021;22:321-31. [Crossref] [PubMed]
17. He Y, Guo P, Tang Y, et al. VISTA3D: A unified segmentation foundation model for 3D medical imaging. In: Proceedings of the Computer Vision and Pattern Recognition Conference. 2025:20863-73.
18. Antonia SJ, Villegas A, Daniel D, et al. Overall Survival with Durvalumab after Chemoradiotherapy in Stage III NSCLC. N Engl J Med 2018;379:2342-50. [Crossref] [PubMed]
19. Lambin P, Leijenaar RTH, Deist TM, et al. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol 2017;14:749-62. [Crossref] [PubMed]
20. Tomita N, Kodaira T, Hida T, et al. The impact of radiation dose and fractionation on outcomes for limited-stage small-cell lung cancer. Int J Radiat Oncol Biol Phys 2010;76:1121-6. [Crossref] [PubMed]
21. Basch E, Deal AM, Dueck AC, et al. Overall Survival Results of a Trial Assessing Patient-Reported Outcomes for Symptom Monitoring During Routine Cancer Treatment. JAMA 2017;318:197-8. [Crossref] [PubMed]
22. O'Connor JP, Rose CJ, Waterton JC, et al. Imaging intratumor heterogeneity: role in therapy response, resistance, and clinical outcome. Clin Cancer Res 2015;21:249-57. [Crossref] [PubMed]
23. Wu J, Tha KK, Xing L, et al. Radiomics and radiogenomics for precision radiotherapy. J Radiat Res 2018;59:i25-31. [Crossref] [PubMed]