Abstract
BACKGROUND: Making precise treatment decisions in esophageal cancer is essential for enhancing patient outcomes and avoiding overtreatment. Traditional approaches relying on special features or shallow learning models often fail to capture the complex, multi-scale patterns embedded in PET/CT imaging data. Recent advances in deep learning provide an opportunity to build more robust, data-driven systems for predictive modeling in oncology. METHODS: We propose a novel deep learning model that integrates convolutional and transformer-based components based on PET/CT data to support treatment decisions for esophageal cancer. The architecture introduces a Convolutional Feature Extractor with split-based residual blocks for efficient local feature capture, a Multi-scale Pooling module for spatial context aggregation, and an Multilayer Perceptron block for predicting. The model was evaluated using several performance metrics such as AUCROC, F1 score, Balanced Accuracy and benchmarked against state-of-the-art convolutional and transformer backbones such as ConvNeXt and Vision Transformer. RESULTS: The proposed model achieved superior performance across all evaluation metrics, including an AUCROC of 0.9935 and a Balanced Accuracy of 0.9630, outperforming existing models. These results validate the effectiveness of combining local-global representation learning through custom-designed modules. In addition, we conducted ablation studies to further demonstrate the individual contributions and effectiveness of each component within the proposed architecture. By systematically removing or replacing specific modules such as the Convolutional Feature Extractor and Multi-scale Pooling, we observed consistent performance degradation, which highlights the necessity and complementary roles of these components in achieving optimal predictive accuracy. CONCLUSIONS: This study presents a novel hybrid deep learning architecture that enhances treatment decision support for esophageal cancer by leveraging multi-scale spatial encoding. The empirical evidence demonstrates that tailored architectural innovations significantly improve predictive accuracy over existing methods.