Multi-Task Residual Cross-Attention Network for Tumor Segmentation And Lymph Node Metastasis Prediction In Cervical Cancer

Tumor localization and lymph node metastasis (LNM) diagnosis are two important tasks for gynecologist to make decisions in cervical cancer treatments. Aiming to develop an accurate and convenient diagnosis system, we propose a multi-task residual cross-attention network named MRCNet for tumor segmentation and LNM prediction. Specifically, we tackle task correlation with underlying related supervision information, and capture multi-level features by multi-scale convolutional neural network, which equipped with cross-attention module concerning spatial and channel dimensions to emphasize meaningful features. A total of 1123 cervical cancer patients from 13 centers in China are collected to assess the architecture, 2 centers of them were set as an external testing cohort. The experimental results demonstrate the promising inference performance and generalization ability of our MRCNet for both segmentation and classification tasks, which can help doctors make judgments about treatment measures.