Enhancing Tumor Classification in Testicular Cancer: Segmentation-based Pretraining and Multimodal Prediction

抄録

Testicular cancer that metastasizes to retroperitoneal lymph nodes is typically treated with chemotherapy, followed by post-chemotherapy retroperitoneal lymph node dissection (PC-RPLND). A significant concern is that approximately 50% of patients undergoing PC-RPLND have necrotic tissue in the resected specimens, indicating potential overtreatment. In this study, we propose a U-Net-based classification model to distinguish between necrosis and residual teratoma prior to surgery, aiming to reduce unnecessary procedures. The U-Net-based classifier achieves an area under the curve (AUC) of 0.856 and demonstrates superior performance compared to a ResNet50 classifier when results are shown in scatterplots with the results given by Logistic Regression using clinical variables. These plots highlight that the U-Net-based model more accurately identifies benign tissues, supporting clinical decision-making and potentially minimizing unnecessary surgeries in testicular cancer patients.