Spatial associations between esophageal lesions and surrounding tissues in esophageal fistula: a CAM-guided radiomics study.

Objective: Esophageal fistula is a life-threatening complication following treatment of esophageal squamous cell carcinoma (ESCC). Early identification of high-risk patients remains challenging due to the limited predictive value of conventional imaging markers. This study aimed to develop and validate a Grad-CAM-guided deep learning-radiomics framework for risk prediction and spatial biomarker discovery using pre-treatment CT. Methods: In this retrospective study, 73 ESCC patients treated between January 2019 and December 2022 were included, of whom 25 (34.2%) developed esophageal fistula within one year. A 3D convolutional neural network (3D-CNN) was trained using stratified five-fold cross-validation. Grad-CAM was applied to localize discriminative regions, guiding radiomics feature extraction from the bilateral lungs, thoracic spine, and esophageal tumor. Radscores were constructed using least absolute shrinkage and selection operator regression. Logistic regression analyses were performed to identify independent predictors. Results: The 3D-CNN demonstrated stable performance across folds, achieving mean accuracy, sensitivity, and AUC of 0.809, 0.873, and 0.848, respectively. Grad-CAM revealed prominent activation differences in the lungs and thoracic spine rather than the central mediastinum. Radiomics analysis confirmed significant textural differences in all three regions (P < 0.001), with higher Radscores observed in the fistula group. In conventional multivariable logistic regression, Lung Radscore (OR = 11.55, P = 0.038) and Esophageal Tumor Radscore (OR = 192.3, P = 0.040) were associated with esophageal fistula, whereas conventional CT parameters lost statistical significance. Firth penalized logistic regression yielded more conservative estimates, with Esophageal Tumor Radscore remaining significant and Lung Radscore showing a borderline association. Conclusion: The Grad-CAM-guided 3D-CNN radiomics framework identified tumor-related and extratumoral imaging patterns associated with esophageal fistula, providing exploratory spatial biomarkers that warrant further validation.