3642 - Deep Learning for Lymph Node Station Metastatic Classification in Esophageal Cancer via Region-Aware Mixture-of-Experts Modeling

Purpose/Objective(s): Assessment of lymph node (LN) metastasis in computed tomography (CT) is a clinically important task for esophageal cancer treatment and treatment planning. Deep learning would have the potential to address this issue by learning from large- scale accurately labeled data if available. However, from the surgical procedure in LN dissection, pathology report only indicates the number of dissected LNs in each lymph node station (LN-station) and the number of metastatic ones found in the respective LN-station. As a result, it is very difficult to establish the one-to-one matching between LN instances observed in CT and their metastasis conditions indicated in pathology report. In contrast, metastasis labels on LN-station are readily retrieved from the pathology reports at scale. Hence, in this work, we directly classify LN-station metastasis using a region-aware mixture-of-experts deep modeling.

Materials/Methods: We collected a dataset consisting of 1153 esophageal cancer patients who had a preoperative contrast-enhanced CT and underwent esophagectomy treatment with detailed pathology report. We cropped each LN-station using a 96×96×32 ROI as the input. To solve the LN-station classification task, we introduce a new LN station-aware mixture-of-experts deep learning model, where the expert is trained to specialize on learning the metastasis features in a group of LN-station that partitioned based on the anatomy. We further incorporate a LN prior attention loss to explicitly regularize the deep network to focus on LN instances inside the LN-station. To evaluate model performances, we conduct five-fold cross-validation by splitting the data at patient level.

Results: Table 1 summarizes the LN-station metastatic classification performance. We compare with the transformer classification network MobileVitv2 that serves as the backbone of our model. It is observed that incorporating the proposed mixture-of-experts module significantly improves the performance by 3.56% AUC, 6.71% Specificity at 80% recall and 7.08% recall at 80% specificity. MobileVitv2 is used as the basic classification network in all comparison methods. S and R represent specificity and recall (sensitivity), respectively.

Conclusion: We developed a LN station-aware mixture-of-experts (MoE) deep learning model for LN station metastatic classification using CT scans of 1153 esophageal cancer patients. Our findings suggest that the proposed model substantially enhances the LN-station classification performance and has the potential to help with cancer treatment and planning.