Tumor-lymph cross-plane projection reveals spatial relationship features: a ResNet-CBAM model for prognostic prediction in esophageal cancer.

Background: Prognostic models for esophageal cancer based on contrast-enhanced chest CT can aid thoracic surgeons in developing personalized treatment plans to optimize patient outcomes. However, the extensive lymphatic drainage and early lymph node metastasis of the esophagus present significant challenges in extracting and analyzing meaningful lymph node characteristics. Previous studies have primarily focused on tumor and lymph node features separately, overlooking spatial correlations such as position, direction, and volumetric ratio. Methods: A total of 285 patients who underwent radical resection surgery at Fujian Provincial Hospital from 2018 to 2022 were retrospectively analyzed. This study introduced a tumor-lymph node projection plane, created by projecting lymph node ROIs onto the tumor ROI plane. A ResNet-CBAM model, integrating a residual convolutional neural network with a CBAM attention module, was employed for feature extraction and survival prediction. The PJ group utilized tumor-lymph node projection planes as training data, while the TM and ZC groups utilized tumor ROIs and concatenated images of tumor and lymph node ROIs, respectively, as controls. Additional comparisons were made with traditional machine learning models (support vector machines, logistic regression, and K-nearest neighbors). Survival outcomes (median, 1-year, 3-year, 5-year) were used as target labels to evaluate model performance in distinguishing high-risk patients and predicting both short- and long-term survival. Results: In the PJ group, the ResNet-CBAM model achieved accuracy rates of 0.766, 0.981, 0.883, and 0.778 for predicting median, 1-year, 3-year, and 5-year survival, respectively. Its corresponding AUC values for 1-, 3-, and 5-year survival were 0.992, 0.913, and 0.835. Kaplan-Meier survival analysis revealed significant differences between high- and low-risk groups identified by the model. The ResNet-CBAM model outperformed those in the TM and ZC groups in distinguishing high-risk patients and predicting both short- and long-term survival. Compared to machine learning models, it demonstrated superior performance in long-term survival prediction. Conclusion: The ResNet-CBAM model trained on tumor-lymph projection planes effectively distinguished high-risk esophageal cancer patients and outperformed traditional models in predicting survival outcomes. By capturing spatial relationships between tumors and lymph nodes, it demonstrated enhanced predictive efficiency.