Interpretable deep learning and radiomics for predicting the occurrence of brain metastases following prophylactic brain irradiation in small cell lung cancer

Introduction

Small cell lung cancer (SCLC), a highly aggressive malignancy, accounts for approximately 15% of all lung cancers. It progresses rapidly and is prone to distant metastasis even in the early stages (1,2). Brain metastasis is a common and serious complication in SCLC patients and a key factor affecting prognosis and quality of life (3,4). Previous studies have shown that the risk of brain metastasis in small-cell lung cancer is closely associated with the biological characteristics of the primary tumor. In particular, RB1 alterations in primary tumor tissue have been associated with an increased risk of brain metastasis (5). In addition, the morphology and spatial heterogeneity of the primary tumor reflected on pretreatment chest computed tomography (CT) may, to some extent, represent tumor aggressiveness and the potential for distant metastasis. CT radiomics features extracted from primary SCLC lesions have also shown value in stratifying the risk of brain metastasis (6). Therefore, the imaging phenotype of the primary tumor may serve as a potential noninvasive biomarker for assessing the risk of brain metastasis after prophylactic cranial irradiation (PCI). To reduce the incidence of brain metastasis, PCI has become the standard treatment strategy for patients with limited-stage and some extensive-stage SCLC (7,8). However, clinical practice indicates that even after receiving PCI, a significant proportion of patients still develops brain metastases post-treatment, and the mechanisms and risk factors underlying this phenomenon remain incompletely understood (9,10). Therefore, predicting whether brain metastases will occur in SCLC patients following PCI is of great significance for making personalized treatment decisions, optimizing follow-up strategies, and improving patient prognosis.

Traditional prediction models based on clinical features currently exhibit limited discriminatory power, making it challenging to meet the demands of precision medicine (11). With the advancement of artificial intelligence in medical imaging, radiomics offers a new approach to characterizing tumor phenotypes by extracting quantitative image features through high-throughput analysis, and has shown promise in predicting brain metastases in SCLC (12,13). However, traditionally hand-engineered radiomics features often rely on prior knowledge and have limitations in representing complex, high-dimensional imaging patterns (14). Furthermore, while existing studies have primarily focused on predicting the risk of brain metastasis at initial diagnosis or during treatment in SCLC patients, there remains a lack of in-depth research on brain metastasis prediction specifically for this subgroup following PCI. In recent years, artificial intelligence technologies, particularly convolutional neural networks (CNNs), have demonstrated strong performance in the image analysis of various cancers by automatically learning multi-level abstract features from images (15). Specifically, CT-based deep learning has proven to significantly outperform clinical models alone in predicting overall survival in SCLC patients and identifying those who benefit from PCI (16,17). In the field of medical imaging, two-dimensional (2D) deep learning models struggle to fully utilize three-dimensional (3D) spatial context information, while 3D deep learning models, although capable of capturing volumetric information, require substantial data volumes and computational resources and are prone to overfitting in small-sample medical datasets (18). Two-and-a-half-dimensional (2.5D) multi-channel deep learning extracts and fuses features from views reconstructed across multiple planes, thereby capturing spatial information from multiple directions while relatively alleviating data and computational constraints (19). This method has been successfully applied to the differentiation of pathological subtypes of brain metastases and has been shown to effectively integrate local and global features to improve model performance in other medical image classification tasks (20).

Machine learning models demonstrate outstanding performance, but their decision-making processes are often considered to be ‘black boxes’ that lack interpretability (21,22). In high-risk fields such as healthcare, understanding the basis for a model’s decisions is crucial for ensuring its ethical and reliable operation (23). Therefore, it is essential to integrate explainable artificial intelligence (XAI) techniques, such as Gradient-Weighted Class Activation Maps (Grad-CAM) and Shapley additive explanations (SHAP), when building predictive models (24). These methods can visualize key image regions and quantify feature contributions, which is vital for enhancing clinicians’ trust in the models and verifying that their decisions are based on reasonable radiological features (25,26).

Against this backdrop, this study developed and validated an interpretable fusion prediction model based on 2.5D multi-channel deep learning features derived from CT images, clinical features, and traditional radiomics features to predict the risk of brain metastasis in SCLC patients following PCI. Concurrently, in-depth analysis of the model’s decision-making basis using interpretability techniques enhances the model’s interpretability and clinical credibility. The establishment of this model is expected to assist clinicians in identifying high-risk groups for brain metastasis following PCI, optimizing follow-up frequency and timing of intervention, avoiding unnecessary prophylactic irradiation, or enabling early intervention, ultimately improving patient prognosis and quality of life. We present this article in accordance with the TRIPOD reporting checklist (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-0708/rc).

Methods

Patients

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Shanghai Chest Hospital, Shanghai Jiao Tong University School of Medicine (No. KS24019). This retrospective study used de-identified patient data obtained from medical records. All records were anonymized prior to analysis to ensure patient confidentiality, and since the data could not be linked back to individual patients, the requirement for informed consent was waived by the Ethics Committee. The study design and protocol are shown in Figure 1. A retrospective analysis was conducted on 185 patients with SCLC admitted to Shanghai Chest Hospital, Shanghai Jiao Tong University School of Medicine, between 2008 and 2017. As shown in Figure 2, the inclusion criteria were as follows: (I) patients with histopathologically confirmed SCLC; (II) Chest CT images acquired before the initiation of anticancer treatment were available, with the primary pulmonary tumor clearly visible and suitable for region of interest (ROI) delineation; (III) no evidence of brain metastases on baseline brain magnetic resonance imaging (MRI)/CT prior to treatment; (IV) completion of the prescribed prophylactic whole-brain irradiation regimen; and (V) availability of complete baseline clinical data, follow-up imaging data, and clear records of disease outcome. Exclusion criteria included: (I) radiologically confirmed brain metastases prior to PCI treatment; (II) history of other malignancies; (III) failure to complete the full course of PCI treatment; (IV) incomplete clinical or imaging follow-up data, making it impossible to clearly determine the status of brain metastases. Ultimately, this study included 185 patients, who were randomly assigned to a training group (N=148) and a test group (N=37).

Figure 1 The study’s flowchart. 2.5D, two-and-a-half-dimensional; 2D, two-dimensional; 3D, three-dimensional; AUC, area under the curve; BM, brain metastasis; CI, confidence interval; DL, deep learning; Rad, radiomics; ROI, region of interest; SCLC, small cell lung cancer.

Figure 2 Cohort diagram for the study of a predictive model for brain metastasis in SCLC patients. BM, brain metastasis; CT, computed tomography; LC, lung cancer; MRI, magnetic resonance imaging; PCI, prophylactic cranial irradiation; SCLC, small cell lung cancer.

Between 2008 and 2017, patients with limited-stage SCLC generally received induction chemotherapy followed by concurrent chemoradiotherapy. Sequential chemoradiotherapy was used for patients who could not tolerate concurrent treatment. Thoracic radiotherapy was delivered using intensity-modulated radiation therapy at a prescribed dose of 60 Gy in 30 fractions. Patients with extensive-stage disease received chemotherapy, and thoracic radiotherapy was selectively administered according to the patient’s clinical condition at a dose of 30 Gy in 10 fractions or 40 to 50 Gy in 20 to 25 fractions. All patients received first-line chemotherapy consisting of etoposide combined with either cisplatin or carboplatin. Treatment was administered every 21 days for a total of 4 to 6 cycles. PCI was offered to patients whose thoracic lesions achieved a complete or partial response after treatment, who had no evidence of brain metastasis on baseline brain imaging, who had no contraindications to cranial irradiation, and who voluntarily agreed to receive PCI. The indications for PCI and the principal treatment strategies remained largely unchanged during the study period, and approximately 40% to 50% of patients with SCLC received PCI. All PCI procedures used conventional whole-brain irradiation without hippocampal avoidance at a prescribed dose of 25 Gy in 10 fractions, which was completed over approximately 14 days.

All patients were followed from the completion of PCI until May 2023. Brain metastasis was primarily assessed using contrast-enhanced brain MRI. Contrast-enhanced brain CT was used when MRI could not be performed. More than 90% of the patients in this study underwent follow-up with contrast-enhanced brain MRI. Brain metastasis was determined on the basis of imaging findings regardless of the presence or absence of neurological symptoms. Brain MRI was generally performed every 3 to 4 months during the first year after PCI and every 6 months during the second year. Subsequent examinations were scheduled according to clinical needs (27). Patients who did not develop brain metastasis were required to complete at least 24 months of brain imaging follow-up, with no evidence of brain metastasis on the final examination. Patients with insufficient follow-up or an indeterminate outcome were excluded (28).

Image acquisition

Pretreatment chest CT images of patients with small-cell lung cancer were retrospectively collected for radiomics analysis. The primary pulmonary tumor identified on chest CT was defined as the ROI. CT images were acquired at a tube voltage of 120 kV and a tube current of 250 mA. The image matrix was 512×512, the in-plane pixel spacing was 0.98 mm × 0.98 mm, and the original slice thickness was 5 or 6 mm. All CT images were exported in Digital Imaging and Communications in Medicine format and subsequently converted to Neuroimaging Informatics Technology Initiative format with the .nii file extension. Before region-of-interest segmentation and radiomics feature extraction, all images were resampled to a uniform voxel spacing of 1 mm × 1 mm × 6 mm. This procedure was performed to improve spatial consistency and reduce the influence of differences in voxel dimensions on the extracted radiomics features. Image intensity normalization was then applied to minimize variations in gray-level intensity among images. The primary pulmonary tumors were manually delineated on each axial slice using ITK-SNAP by oncologists with extensive clinical experience. Cases with discrepancies in tumor delineation were reviewed by a senior physician with more than 20 years of clinical experience, who made the final decision.

Radiomics procedure

Feature extraction

Radiomics features were systematically extracted from each patient’s CT images using PyRadiomics (29). A total of 1,561 quantitative features were obtained, covering seven major feature categories: 374 gray-level co-occurrence matrix (GLCM) features, 272 gray-level run-length matrix (GLRLM) features, 272 gray-level scale-zone matrix (GLSZM) features, 306 first-order statistical features, 238 gray-level dependency matrix (GLDM) features, 85 neighborhood gray-level tone difference matrix (NGTDM) features, and 14 morphological features.

Feature selection

In this study, all extracted features were standardized using Z-scores, and their statistical significance was assessed using t-tests and Mann-Whitney U tests; only features with P<0.05 were retained. To mitigate multicollinearity, we analyzed the correlations among features using Pearson’s correlation coefficient and excluded highly redundant features with a correlation coefficient ≥0.9. Within a 10-fold cross-validation framework, we further optimized the feature set using least absolute shrinkage and selection operator (LASSO) regression to determine the optimal regularization parameter λ (30).

Deep learning procedure

2D image cropping

From all axial CT slices containing the primary lung tumor, select the slice with the largest cross-sectional area of the tumor ROI as the representative image for each patient. The slice containing the ROI is selected and cropped to its smallest bounding box to reduce computational complexity and minimize background noise. The image is normalized using Z-scaling to standardize the brightness across all RGB channels before being fed into the model. During training, real-time data augmentation is applied using random cropping and horizontal and vertical flipping; during testing, only normalization is performed.

2D model training

This study employs a ResNet101 network pre-trained on the ImageNet dataset, which contains millions of annotated images spanning thousands of categories. Through a transfer learning strategy, we leverage ImageNet’s powerful feature extraction capabilities on the training set and perform global fine-tuning of the network parameters to adapt ResNet101 to the brain transfer prediction task (31).

2.5D image cropping

To effectively utilize 2.5D data, we adopted a multi-view ROI extraction strategy (32,33). For each 3D volumetric dataset, we centered on the target anatomical structure and extracted ROI in orthogonal planes (axial, sagittal, and coronal). We then stacked these 2D slices along the channel dimension to form a multi-channel input.

2.5D model training

This study used an ImageNet-pretrained ResNet101 model for transfer learning (34). The input layer was modified to accommodate multichannel data, and deep features extracted from the network were fused before risk prediction was generated through a fully connected layer. The model was trained using cross-entropy loss for binary classification. Optimization was performed using a cosine annealing learning rate schedule, a regularized stochastic gradient descent optimizer, learning rate warm-up, fine-tuning, and early stopping (35).

3D image cropping

The volume containing the maximum ROI was selected as the representative image. To reduce the complexity of the analysis and suppress background noise, the image is cropped to the minimum bounding box (MBB). After Z-score normalization, the image is used as the model input. During training, real-time data augmentation techniques such as random cropping and horizontal and vertical flipping are applied; during testing, only normalization is performed.

3D model training

This study investigates the performance of ResNet101 in enhancing the performance of traditional CNN models. During the model training phase, we resample each cropped sub-region image to a fixed size of 64×64×48 using nearest-neighbor interpolation, and use these as the model’s input. To improve the model’s generalization ability, we fine-tune the learning rate by adopting a cosine-decay learning rate scheduling strategy, defined as follows:

ηt=ηmini+12(ηmaxi−ηmini)(1+cos(TcurTiπ))

Specifically, ηmini=0 represents the minimum learning rate, ηmaxi=0.01 represents the maximum learning rate, Tcur denotes the current epoch number, and Ti=30 indicates the total number of epochs in the training process. Other key hyperparameters include: stochastic gradient descent as the optimizer and softmax cross-entropy as the loss function.

Signature construction

After the training and test cohorts were defined, all model development procedures were performed exclusively within the training cohort. The CNN model and its hyperparameters were determined through internal cross-validation, and deep learning features were extracted from the penultimate layer of the final model. Feature standardization and principal component analysis were performed using only the training data. Principal components accounting for 95% of the cumulative explained variance were retained, and the resulting transformation parameters were then fixed and applied to the test cohort. A pre-fusion strategy was subsequently used to integrate the deep learning features, radiomics features, and selected clinical features. LASSO-based feature selection, model selection, and hyperparameter optimization were all completed within the training cohort. The final model was then applied to the independent test cohort without further adjustment for performance evaluation.

Metrics

The primary endpoint of this study was whether patients developed brain metastases during the follow-up period after prophylactic brain irradiation. By comparing the model-predicted probability of brain metastases after PCI with the actual observed binary outcome, receiver operating characteristic (ROC) curves were plotted and the area under the curve (AUC) was calculated for both the training and testing cohorts. Model calibration was analyzed using calibration curves, and its reliability was verified using the Hosmer-Lemeshow goodness-of-fit test. Additionally, decision curve analysis (DCA) was employed to evaluate the clinical utility of the predictive model, thereby providing a deeper understanding of its potential benefits in clinical settings.

Grad-CAM and SHAP interpretation

Grad-CAM was applied to the final convolutional layer of the last residual module in ResNet101 to spatially localize decision-critical regions through gradient-weighted activation mapping. The optimal fusion model is quantitatively interpreted using SHAP. The summary plot is sorted in descending order of global feature importance, with distribution density displayed as a horizontal scatter plot stacked vertically. Feature values are shaded from blue to red in ascending order, simultaneously illustrating the range of contributions, distribution patterns, and the correlation between feature values and their impact.

Statistical analyses

All analyses were performed using Python 3.7.12. Statistical analyses were conducted using statsmodels 0.13.2, and radiomics features were extracted using PyRadiomics 3.0.1. Machine learning models were implemented using scikit-learn 1.0.2, while deep learning models were developed using PyTorch 1.11.0 with CUDA 11.3 and cuDNN 8.2.1.

Results

Clinical features

The normality of clinical characteristics was verified using the Shapiro-Wilk test. Depending on the distribution of continuous variables, analyses were performed using either a t-test or the Mann-Whitney U test. Categorical variables were analyzed using the chi-square (χ²) test. Baseline characteristics for all cohorts are detailed in Table 1. It is worth noting that p-values greater than 0.05 between cohorts indicate no significant differences, thereby confirming the unbiased nature of the group stratification.

Systematic univariate and multivariate logistic regression analyses were performed on all clinical characteristics to assess the association between each variable and the risk of brain metastasis. In the univariate analysis, a logistic regression model was used to calculate the odds ratio (OR) and 95% confidence interval (CI) for each variable, with a significance level set at P<0.05. Based on the results of the univariate analysis, statistically significant characteristics were included in the multivariate analysis, and indicators that remained significant were ultimately selected to construct a comprehensive predictive model. Detailed results are presented in Table 2.

2D_DL, 2.5D_DL multi-channel deep learning, 3D_DL feature selection

The 2D_clinic_Rad fusion model includes 2D deep learning features from CT images, radiomics features, and clinical features. The 2.5D_clinic_Rad fusion model includes 2.5D multi-channel deep learning features from CT images, traditional omics features, and clinical features. The 3D_clinic_Rad fusion model includes 3D deep learning features, radiomics features, and clinical features. After calculating the standard correlation coefficients, the three models retained 313, 369, and 337 features, respectively. Feature selection for the three models was performed using LASSO, with the results shown in Figure 3.

Figure 3 Feature selection. (A-I) Feature selection using the LASSO regression model for different models. 2.5d, two-and-a-half-dimensional; 2d, two-dimensional; 3d, three-dimensional; DL, deep learning; LASSO, least absolute shrinkage and selection operator; MSE, mean squared error; Rad, radiomics.

Performance evaluation and model comparison

Compared to other models in the training and test sets, the 2.5D_clinic_Rad multi-channel model demonstrated the highest predictive performance, with an AUC of 0.841 in the test set. Among the other models, the AUC values on the test set were 0.750 for the clinical model, 0.691 for the Rad model, 0.703 for the 2D_clinic_Rad model, and 0.762 for the 3D_clinic_Rad model, as shown in Table 3. Overall, the machine learning-based 2.5D_clinic_Rad multi-channel model demonstrated significantly superior discriminatory ability compared to traditional radiomics models, with a test set AUC of 0.841, indicating high discriminatory power and stability. In contrast, the AUC of the traditional radiomics model was only 0.691, suggesting limited discriminatory performance. As shown in Figure 4A,4B, the 2.5D_clinic_Rad multi-channel model performed most notably in this task, demonstrating strong clinical predictive capability.

Figure 4 Comparison of the model’s performance in the training and test cohort. (A,B) Comparison of the AUCs in the training cohort (A) and test cohort (B). Clinic represents the clinical model. Rad represents the radiomics model. (C,D) Comparison of the DCA in the training cohort (C) and test cohort (D). (E,F) Comparison of the calibration curve in the training cohort (E) and test cohort (F). 2.5d, two-and-a-half-dimensional; 2d, two-dimensional; 3d, three-dimensional; AUC, area under the curve; CI, confidence interval; DCA, decision curve analysis; Rad, radiomics.

Based on the results of the DCA, as shown in Figure 4C,4D, the 2.5D_clinic_Rad model maintains the highest net benefit across a wide range of threshold probability intervals, with its curve consistently exceeding those of other models, suggesting that this model offers the greatest clinical utility in the training data; The 2D_clinic_Rad model also exhibits stable positive net benefits across the threshold probability range, whereas the 3D_clinic_Rad model shows positive net benefits only in the low-threshold range, with net benefits subsequently declining rapidly and turning negative. The net benefits of traditional omics and clinical models are relatively limited. Overall, the 2.5D_RandomForest model demonstrates the most stable clinical net benefits.

As shown in the calibration curve results (Figure 4E,4F), the calibration curve of the 2.5D_clinic_Rad multi-channel model most closely follows the ideal calibration line (Perfectly calibrated) in the medium-to-high probability range, with only a slight overestimation in the high-probability segment; The 2D_clinic_Rad model exhibits underestimation in the low-probability range but gradually converges toward the ideal calibration line in the medium-to-high probability ranges; traditional omics and clinical models demonstrate good calibration consistency in the training set, whereas the 3D_clinic_Rad model exhibits a significant overestimation bias across the entire probability range.

Explanation and visualization of 2.5D_clinic_Rad model

The above results indicate that the 2.5D_clinic_Rad multi-channel model developed in this study can effectively utilize preoperative CT scans to predict the risk of brain metastasis following prophylactic brain irradiation in patients with SCLC. As shown in the Grad-CAM visualization in Figure 5, the model can accurately identify key lesion regions. In the heatmaps of both sample groups, the highly activated red regions show a high degree of correspondence with manually annotated lesions. In the first group, these regions are concentrated in the core of the lesions, while in the second group, they cover the peripheral infiltration zones, with activation intensity significantly higher than that of normal tissue, maintaining spatial consistency across multimodal sections. This validates the rationality and interpretability of the model’s feature learning, providing an intuitive basis for clinical decision-making.

Figure 5 Visualization of the 2.5D_clinic_Rad model. (A) Representative sample showing high activation predominantly in the lesion core. (B) Representative sample showing high activation extending to the peripheral infiltration zone. 2.5D, two-and-a-half-dimensional; Rad, radiomics.

SHAP provided a quantitative interpretation of the model. As shown in Figure 6, among all features, DL_49 was the most important feature. High values (red) corresponded to positive SHAP values, driving the model toward higher-risk predictions, while low values (blue) corresponded to negative SHAP values, driving the model toward lower-risk predictions, exerting the most significant influence on the model’s decision-making. Most other radiomics features, such as log_sigma_3_0_mm_3D_firstorder_Maximum and square_glszm_ZoneEntropy, as well as deep learning features such as DL_31 and DL_2, also exhibit similar positive correlations, meaning that higher feature values tend to drive high-risk predictions, whereas Disease_progression shows an inverse correlation: its high values correspond to negative SHAP values, reducing the model’s tendency toward high-risk predictions. Overall, the model’s decision-making is driven jointly by deep learning features and radiomics features. This figure effectively quantifies the differences in contribution among these features, providing intuitive and interpretable insights for the model.

Figure 6 SHAP summary plot of 2.5D_clinic_Rad model. 2.5D, two-and-a-half-dimensional; 3D, three-dimensional; DL, deep learning; LHH, low-high-high; LLL, low-low-low; Rad, radiomics; SHAP, Shapley additive explanations.

Waterfall plots and force plots visually illustrate the contribution of individual patient characteristics to the model’s prediction results. Using the cohort’s average baseline SHAP value E[f(X)] =1.108 as a reference, the final predicted value f(x) =0.894 for the patient in Figure 7A is lower than the baseline, primarily driven by positive contributions from features such as DL_49 (+0.36) and DL_31 (+0.15), among others, combined with the negative suppression from features such as wavelet_LLL_firstorder_90Percentile (−0.23) and square_glszm_ZoneEntropy (−0.21). In Figure 7B, the final predicted value f(x) =1.603 for the patient is higher than the baseline. The core positive contributions come from features such as log_sigma_3_0_mm_3D_firstorder_Maximum (+0.20) and DL_49 (+0.17), with only a few features exerting a slight negative influence; The bar chart below further quantifies the cumulative impact of each feature, visually illustrating the dominant role of features in the prediction of individual samples and providing interpretable insights into the model’s black-box decision-making.

Figure 7 SHAP waterfall plot and force plot of the 2.5D_clinic_Rad model. (A) Representative patient with a predicted value of f(x) =0.89, below the baseline value E[f(X)] =1.108. (B) Representative patient with a predicted value of f(x) =1.60, above the baseline value E[f(X)] =1.108. 2.5D, two-and-a-half-dimensional; 3D, three-dimensional; DL, deep learning; LHH, low-high-high; LLL, low-low-low; Rad, radiomics; SHAP, Shapley additive explanations.

Discussion

This study developed and validated a predictive model that integrates 2.5D multi-channel deep learning features, traditional radiomics features, and clinical features to assess the risk of brain metastasis in patients with SCLC following PCI. First, our proposed 2.5D_clinic_Rad fusion model demonstrated excellent predictive performance, with a test set AUC of 0.841 that was superior to clinical models, radiomics models, and fusion models based on 2D or 3D deep learning features. Second, interpretability analysis validated the rationality of the model’s decision-making process and quantitatively elucidated the significant contributions of deep learning features to the prediction process. These results indicate that the 2.5D_clinic_Rad fusion model can reliably predict the risk of brain metastasis in patients with SCLC following PCI, and the interpretability of its decision-making process further enhances the model’s clinical credibility.

In this study, 2.5D multi-channel deep learning was able to extract better spatial features from images, and the performance of the 2.5D_clinic_Rad model further demonstrated the value of 2.5D multi-channel deep learning in medical image analysis. Compared to 2D deep learning, which can only utilize information from a single plane, 2.5D multi-channel deep learning effectively integrates the 3D spatial context of tumors by simultaneously acquiring views from the axial, coronal, and sagittal planes. This multi-view input enables the model to capture morphological heterogeneity and invasive characteristics of lesions across different directions, features that may be closely related to a tumor’s metastatic potential (36). Additionally, compared to 3D CNN that require processing a complete 3D voxel grid, the 2.5D model adopted in this study significantly reduces computational resource requirements and model parameter complexity by compressing 3D information into multiple 2D channels. For a medium-sized dataset comprising 185 patients, this effectively mitigates the overfitting issues that 3D CNN are prone to in low-sample-size scenarios. Consequently, the 2.5D model achieves an optimal balance between preserving sufficient radiological information and controlling model complexity, which is likely the fundamental reason for its superior performance over 2D and 3D models.

Furthermore, DCA further confirmed the clinical utility of the 2.5D_clinic_Rad model. As shown in Figure 4C,4D, this model maintained the highest net benefit across a wide range of threshold probability intervals, significantly outperforming the other comparison models, indicating that it offers a superior benefit-risk trade-off in clinical decision-making. Although the 2D_clinic_Rad model exhibits positive net benefit in certain intervals, the net benefit of the 3D_clinic_Rad model is effective only in the low-threshold range, rapidly declining thereafter and even turning negative. This may be related to overfitting of the 3D model on small datasets and decision instability caused by its parameter complexity.

Meanwhile, the results of the calibration curve analysis (Figure 4E,4F) show that the predicted probabilities of the 2.5D_clinic_Rad model align most closely with the ideal calibration line in the medium-to-high probability range, with only a slight overestimation in the high-probability segment, reflecting its strong probability calibration capability. The 2D_clinic_Rad model underestimates risk in the low-probability segment, while the 3D_clinic_Rad model exhibits a significant overestimation bias across the entire probability range, further highlighting the optimization challenges associated with complex 3D structures in medium-sized datasets. Therefore, the 2.5D_clinic_Rad model demonstrates good discriminatory power, calibration, and clinical net benefit, and can accurately reflect actual risk. This indicates that the model has the potential to aid clinical decision-making, particularly in scenarios requiring objective quantitative assessment or cross-validation of expert opinions, and may provide valuable reference for radiologists.

The Grad-CAM visualization heatmap shows that the model’s areas of interest closely align with the lesion regions outlined by radiologists, and even highlight the more aggressive portions within the lesions. This provides intuitive evidence that the model does not make black-box predictions, but rather bases its decisions on radiological findings consistent with clinical judgment, thereby enhancing the trust that radiologists and clinicians place in the model.

SHAP analysis reveals that deep learning features are crucial for the model’s predictive capability. Traditional bioinformatics relies on manually predefined feature descriptors, which, while interpretable, may struggle to fully capture the complex microscopic patterns of tumors (37). Deep learning, through hierarchical abstraction, can automatically extract deep image representations from pixels that are potentially associated with prognosis (38). This study integrates deep learning features, radiomics features, and clinical features to predict the risk of brain metastasis in patients with SCLC following SCLC. Furthermore, SHAP analysis quantifies the contribution of each feature at both the global and individual levels. The global analysis identifies key deep learning features such as DL_49 and DL_31. Individualized waterfall plots clearly demonstrate how a specific patient’s prediction is determined by their unique combination of features (39). This interpretability transforms the model output from a simple risk score into a traceable, discussable decision-making process, helping clinicians understand the rationale behind the model’s judgments (40).

The model developed in this study has clinical practical significance. For patients identified by the model as being at high risk of brain metastasis following PCI, clinicians may consider adopting more aggressive intervention strategies, such as increasing the frequency of follow-up imaging, exploring consolidation or prophylactic treatment with new drugs, or optimizing the dose and regimen of the initial PCI (41). DCA indicates that the 2.5D_clinic_Rad model offers higher clinical net benefit across a wide range of threshold probabilities, providing a decision-economic rationale for its practical application. From a biological perspective, the deep learning features underlying the model may implicitly encode intrinsic tumor characteristics associated with a propensity for brain metastasis, such as specific angiogenesis patterns, cellular density heterogeneity, or local microenvironmental interaction features (42). These aspects warrant further exploration and validation in future studies incorporating pathogenomic analysis. In recent years, the use of dose-escalated hyperfractionated thoracic radiotherapy and durvalumab consolidation therapy following concurrent chemoradiotherapy has further improved clinical outcomes in patients with limited-stage small-cell lung cancer (43,44). Despite the continuing evolution of treatment strategies, the imaging phenotype of the primary tumor on pretreatment chest CT may still provide valuable complementary information for stratifying the risk of brain metastasis after PCI (5,45). In the era of MRI surveillance, the survival benefit of PCI remains controversial and may be influenced by multiple clinical factors. In this study, pretreatment chest CT radiomics and deep learning features were integrated with conventional clinical variables. The combined model outperformed the clinical model, suggesting that the imaging phenotype of the primary tumor provides additional value for risk stratification. Clinical factors including age, sex, smoking status, tumor-node-metastasis (TNM) stage, and treatment modality were incorporated into the analysis. However, owing to the limitations of the retrospective data, variables such as performance status, serum biomarkers, and inflammatory indices could not be comprehensively included. We are currently collecting data from patients treated with contemporary therapeutic regimens and plan to incorporate immunotherapy-related variables to further validate and extend the clinical value of the model in future studies.

This study has several limitations. First, it was a single-center retrospective study with a relatively limited sample size, which may have introduced selection bias. Second, patients with limited-stage and extensive-stage SCLC were analyzed together, which may have increased clinical heterogeneity. Stage-specific subgroup validation could not be performed because of the limited sample size. Further evaluation of the generalizability and robustness of the model is therefore required in large-scale multicenter prospective cohorts. In addition, genomic information such as SCLC molecular subtypes was not included. The identified imaging features may partly reflect underlying molecular heterogeneity, and residual confounding related to molecular factors cannot be fully excluded. Finally, the model was primarily based on baseline CT images. Future integration of longitudinal imaging changes during treatment and blood-based biomarkers may help develop a more comprehensive longitudinal prediction model.

Conclusions

This study successfully developed an interpretable predictive model based on 2.5D multi-channel deep learning feature fusion. The model effectively predicts the risk of brain metastasis in SCLC patients following PCI, with performance significantly superior to that of conventional methods. Through interpretability techniques, we revealed the radiological basis of the model’s decisions and the contributions of key features, thereby effectively enhancing its clinical transparency and credibility. This model holds promise as a non-invasive tool to help identify high-risk patients who may develop brain metastases following PCI, thereby optimizing individualized treatment and follow-up strategies.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-0708/rc

Data Sharing Statement: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-0708/dss

Peer Review File: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-0708/prf

Funding: This work was supported by the National Natural Science Foundation of China (Nos. 82573516, 12305353, and 82373226) and the Shanghai Municipal Health Commission Health Industry Clinical Research Special Project (No. 20234Y0018).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-0708/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. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Shanghai Chest Hospital, Shanghai Jiao Tong University School of Medicine (No. KS24019). This retrospective study used de-identified patient data obtained from medical records. All records were anonymized prior to analysis to ensure patient confidentiality, and since the data could not be linked back to individual patients, the requirement for informed consent was waived by the Ethics Committee.

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. Remon J, Aldea M, Besse B, et al. Small cell lung cancer: a slightly less orphan disease after immunotherapy. Ann Oncol 2021;32:698-709. [Crossref] [PubMed]
2. Yuan M, Zhao Y, Arkenau HT, et al. Signal pathways and precision therapy of small-cell lung cancer. Signal Transduct Target Ther 2022;7:187. [Crossref] [PubMed]
3. Baik CS, Chamberlain MC, Chow LQ. Targeted Therapy for Brain Metastases in EGFR-Mutated and ALK-Rearranged Non-Small-Cell Lung Cancer. J Thorac Oncol 2015;10:1268-78. [Crossref] [PubMed]
4. Levy A, Le Péchoux C, Mistry H, et al. Prophylactic Cranial Irradiation for Limited-Stage Small-Cell Lung Cancer Patients: Secondary Findings From the Prospective Randomized Phase 3 CONVERT Trial. J Thorac Oncol 2019;14:294-7. [Crossref] [PubMed]
5. Li L, Liu N, Zhou T, et al. A biomarker exploration in small-cell lung cancer for brain metastases risk and prophylactic cranial irradiation therapy efficacy. Lung Cancer 2024;196:107959. [Crossref] [PubMed]
6. Zhou Y, Xiao L, Yang S, et al. Radiomics-clinical integration guides prophylactic cranial irradiation decisions in limited-stage small cell lung cancer: a brain metastasis risk stratification model. Transl Lung Cancer Res 2025;14:2584-97. [Crossref] [PubMed]
7. Le Péchoux C, Laplanche A, Faivre-Finn C, et al. Clinical neurological outcome and quality of life among patients with limited small-cell cancer treated with two different doses of prophylactic cranial irradiation in the intergroup phase III trial (PCI99-01, EORTC 22003-08004, RTOG 0212 and IFCT 99-01). Ann Oncol 2011;22:1154-63. [Crossref] [PubMed]
8. Lee HI, Noh JM, Lee TH, et al. Refining Patient Selection for Prophylactic Cranial Irradiation in Limited-Stage SCLC: A Multicenter Cohort Study. J Natl Compr Canc Netw 2025;23:371-8. [Crossref] [PubMed]
9. Gaebe K, Erickson AW, Chen S, et al. Brain metastasis burden and management in patients with small cell lung cancer in Canada: a retrospective, population-based cohort study. EClinicalMedicine 2024;77:102871. [Crossref] [PubMed]
10. Li H, Zhao Y, Ma T, et al. Radiotherapy for extensive-stage small-cell lung cancer in the immunotherapy era. Front Immunol 2023;14:1132482. [Crossref] [PubMed]
11. Chu J, Sun N, Hu W, et al. Bayesian hierarchical lasso Cox model: A 9-gene prognostic signature for overall survival in gastric cancer in an Asian population. PLoS One 2022;17:e0266805. [Crossref] [PubMed]
12. Park H, Tseng SC, Sholl LM, et al. Molecular Characterization and Therapeutic Approaches to Small Cell Lung Cancer: Imaging Implications. Radiology 2022;305:512-25. [Crossref] [PubMed]
13. Zhao J, He Y, Yang X, et al. Assessing treatment outcomes of chemoimmunotherapy in extensive-stage small cell lung cancer: an integrated clinical and radiomics approach. J Immunother Cancer 2023;11:e007492. [Crossref] [PubMed]
14. Marasi A, Milesi D, Aquino D, et al. Glioblastoma survival prediction through MRI and clinical data integration with transfer learning. Int J Comput Assist Radiol Surg 2026;21:473-82. [Crossref] [PubMed]
15. de Margerie-Mellon C, Chassagnon G. Artificial intelligence: A critical review of applications for lung nodule and lung cancer. Diagn Interv Imaging 2023;104:11-7. [Crossref] [PubMed]
16. Zheng X, Liu K, Shen N, et al. Predicting overall survival and prophylactic cranial irradiation benefit in small-cell lung cancer with CT-based deep learning: A retrospective multicenter study. Radiother Oncol 2024;195:110221. [Crossref] [PubMed]
17. Chen Q, Liang H, Zhou L, et al. Deep learning of bone metastasis in small cell lung cancer: A large sample-based study. Front Oncol 2023;13:1097897. [Crossref] [PubMed]
18. Nie Y, Cui Q, Li W, et al. MHAGuideNet: a 3D pre-trained guidance model for Alzheimer's Disease diagnosis using 2D multi-planar sMRI images. BMC Med Imaging 2024;24:338. [Crossref] [PubMed]
19. Yu H, Kong D, Mo X, et al. MRI-based 2.5D deep learning and radiomics effectively predicted microvascular invasion and Ki-67 expression in hepatocellular carcinoma. PLoS One 2025;20:e0336579. [Crossref] [PubMed]
20. Cao J, Liu Y, Feng T, et al. Classifying brain metastases originating from different pathological subtypes of lung cancer via a multimodal magnetic resonance imaging-based deep learning approach. J Thorac Dis 2025;17:5250-9. [Crossref] [PubMed]
21. Long X, Gan F, Fan H, et al. EfficientNetB0-Based End-to-End Diagnostic System for Diabetic Retinopathy Grading and Macular Edema Detection. Diabetes Metab Syndr Obes 2025;18:1311-21. [Crossref] [PubMed]
22. Oh M, Zhang L. DeepGeni: deep generalized interpretable autoencoder elucidates gut microbiota for better cancer immunotherapy. Sci Rep 2023;13:4599. [Crossref] [PubMed]
23. Osadebey M, Liu Q, Fuster-Garcia E, et al. Interpreting deep learning models for glioma survival classification using visualization and textual explanations. BMC Med Inform Decis Mak 2023;23:225. [Crossref] [PubMed]
24. Yang Y, Mohammadzadeh F, Khishe M, et al. Deep learning for sports motion recognition with a high-precision framework for performance enhancement. Sci Rep 2025;15:38861. [Crossref] [PubMed]
25. Xiang H, Xiao Y, Li F, et al. Development and validation of an interpretable model integrating multimodal information for improving ovarian cancer diagnosis. Nat Commun 2024;15:2681. [Crossref] [PubMed]
26. Liu Y, Duffy K, Dy JG, et al. Explainable deep learning for insights in El Niño and river flows. Nat Commun 2023;14:339. [Crossref] [PubMed]
27. Hou Q, Sun B, Yao N, et al. Construction of Brain Metastasis Prediction Model and Optimization of Prophylactic Cranial Irradiation Selection for Limited-Stage Small-Cell Lung Cancer. Cancers (Basel) 2022;14:4906. [Crossref] [PubMed]
28. Arriagada R, Le Chevalier T, Rivière A, et al. Patterns of failure after prophylactic cranial irradiation in small-cell lung cancer: analysis of 505 randomized patients. Ann Oncol 2002;13:748-54. [Crossref] [PubMed]
29. van Griethuysen JJM, Fedorov A, Parmar C, et al. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res 2017;77:e104-7. [Crossref] [PubMed]
30. Xie Y, Shi H, Han B. Bioinformatic analysis of underlying mechanisms of Kawasaki disease via Weighted Gene Correlation Network Analysis (WGCNA) and the Least Absolute Shrinkage and Selection Operator method (LASSO) regression model. BMC Pediatr 2023;23:90. [Crossref] [PubMed]
31. Chang B, Geng Z, Mei J, et al. Application of multimodal deep learning and multi-instance learning fusion techniques in predicting STN-DBS outcomes for Parkinson's disease patients. Neurotherapeutics 2024;21:e00471. [Crossref] [PubMed]
32. Kim SM, Choi MH, Han SH, et al. Comparison of 2D, 2.5D, and 3D landmark localization networks for 3D cephalometry in CT images. BMC Oral Health 2025;25:1843. [Crossref] [PubMed]
33. Jin K, Sun C, Xia L. Development of a hybrid 2.5D deep learning model for glioma survival prediction using T1-weighted MRI from the CGGA database. Eur J Radiol Open 2025;15:100697. [Crossref] [PubMed]
34. Ponzio F, Descombes X, Ambrosetti D. Improving CNNs classification with pathologist-based expertise: the renal cell carcinoma case study. Sci Rep 2023;13:15887. [Crossref] [PubMed]
35. Wang L, Feng L, Zhang J, et al. DCS-NET: a multi-task model for uterine ROI detection and automatic staging of early endometrial cancer in MRI. Sci Rep 2025;15:41024. [Crossref] [PubMed]
36. Huang Z, Qiu Z, Chen S, et al. Optimizing Deep Learning Models for Luminal and Nonluminal Breast Cancer Classification Using Multidimensional ROI in DCE-MRI-A Multicenter Study. Cancer Med 2025;14:e70931. [Crossref] [PubMed]
37. Wu Y, Xu H, Gu Z. Diagnostic performance of artificial intelligence models in predicting liver metastasis in pancreatic ductal adenocarcinoma (PDAC): a systematic review and meta-analysis. Int J Surg 2025; [Epub ahead of print]. [Crossref] [PubMed]
38. Goyal A, Lakhwani K. Integrating advanced deep learning techniques for enhanced detection and classification of citrus leaf and fruit diseases. Sci Rep 2025;15:12659. [Crossref] [PubMed]
39. Du B, Song Z, Ren Z, et al. The global epidemiology, risk factors, and mortality prediction of nocardiosis: an easily missed opportunistic infection. Sci Rep 2025;15:42090. [Crossref] [PubMed]
40. Sun Q, Zhang K, Xu Y, et al. Explainable machine learning for predicting clinical outcomes in HIV/TB co-infection: a comparative retrospective study. BMC Infect Dis 2025;25:1589. [Crossref] [PubMed]
41. Belluomini L, Calvetti L, Inno A, et al. SCLC Treatment in the Immuno-Oncology Era: Current Evidence and Unmet Needs. Front Oncol 2022;12:840783. [Crossref] [PubMed]
42. Bao X, Wu J, Jiang J, et al. Quantitative protein expression of blood-brain barrier transporters in the vasculature of brain metastases of patients with lung and breast cancer. Clin Transl Sci 2021;14:1265-71. [Crossref] [PubMed]
43. 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. Erratum in: Lancet Respir Med 2024 Dec;12(12):e78. [Crossref] [PubMed]
44. Cheng Y, Spigel DR, Cho BC, et al. Durvalumab after Chemoradiotherapy in Limited-Stage Small-Cell Lung Cancer. N Engl J Med 2024;391:1313-27. [Crossref] [PubMed]
45. Zhou G, Zhang Z, Yu P, et al. Predictive value of clinical characteristics on risk and prognosis of synchronous brain metastases in small-cell lung cancer patients: A population-based study. Cancer Med 2023;12:1195-203. [Crossref] [PubMed]