To develop and validate a Grad-CAM–guided deep learning–radiomics framework for risk prediction and spatial biomarker discovery using pre-treatment CT in patients with esophageal squamous cell carcinoma (ESCC).
Approach:
Study Design: A retrospective study involving 73 ESCC patients treated between January 2019 and December 2022, with a focus on those who developed esophageal fistula.
Model Development: A 3D convolutional neural network (3D-CNN) was trained using stratified five-fold cross-validation, with Grad-CAM applied for feature localization.
Statistical Analysis: Logistic regression analyses were performed to identify independent predictors of esophageal fistula.
Key Findings:
The 3D-CNN achieved mean accuracy of 0.809, sensitivity of 0.873, and AUC of 0.848.
Grad-CAM indicated significant activation differences in the lungs and thoracic spine.
Higher Radscores were observed in the fistula group, with significant textural differences in the analyzed regions (P < 0.001).
Lung Radscore (OR = 11.55, P = 0.038) and Esophageal Tumor Radscore (OR = 192.3, P = 0.040) were significant predictors of esophageal fistula.
Interpretation:
The Grad-CAM–guided 3D-CNN radiomics framework identified tumor-related and extratumoral imaging patterns associated with esophageal fistula, suggesting potential spatial biomarkers for further validation.
Limitations:
The study is retrospective and may be subject to selection bias.
The findings require validation in larger, independent cohorts.
Conclusion:
The study presents a novel approach combining deep learning and radiomics to enhance risk prediction for esophageal fistula in ESCC patients.
