A 3D end-to-end multi-task learning network for predicting lymph node metastasis at multiple nodal stations in gastric cancer

Gastric cancer remains a global health concern with high incidence and mortality rates. Accurate preoperative prediction of lymph node (LN) metastasis is crucial for staging, treatment planning, and prognosis. This study introduces a novel 3D end-to-end lymph node metastasis multi-task learning network (LMML-net) designed to predict LN metastasis across multiple nodal stations in gastric cancer. We analyzed a cohort of 293 patients who underwent gastrectomy with LN dissection. Preoperative CT scans, conducted within two weeks before surgery, were utilized. The LMML-net integrates tumor segmentation and LN metastasis prediction, employing a 3D attention-unet for tumor segmentation and a multi-task learning approach to address metastasis at different nodal stations. LMML-net demonstrated robust predictive performance, achieving AUCs of 0.813, 0.820, and 0.805 for total LN metastasis in training, testing, and validating cohorts, respectively. Notably, the model effectively addressed challenges posed by early gastric cancer and exhibited satisfactory results across various nodal stations. Visualization through GradCam highlighted significant contributions of both tumor and connective tissue areas to the predictions, enhancing the model's interpretability. The LMML-net exhibits strong predictive capabilities for LN metastasis across multiple stations in gastric cancer, including cases of early-stage disease. This innovative approach holds promise for guiding personalized preoperative treatments and surgical planning, potentially improving patient outcomes in gastric cancer management. Code and models will be available at: https://github.com/yangzhi028/LMML-net.