Mutual risk prompt learning with multi-objective optimization for collaborative tumor and peritumor segmentation

Abstract

Early radical surgery, radiotherapy, and other treatments may offer curative effects for tumors. However, the proximity of the tumor to surrounding organs-at-risk (OARs) significantly influences both the surgical outcome and prognosis. For benign tumors, the risk is primarily associated with the tumor's boundaries. In contrast, for malignant tumors, the main challenge lies in balancing the preservation of surrounding organ function while minimizing the risk of tumor recurrence. Therefore, understanding the tumor's characteristics and its anatomical relationships with OARs are essential. Most of the existing studies neglect the constrained interrelations and the potential optimization conflicts between tumor and OARs and easily introduce risks and uncertainties in tumor treatment and OARs protection. Here, we propose a novel multi-objective segmentation network for tumor and OARs, called ROJS-Net, which incorporates mutual risk prompt learning and multi-gate mixture of experts to achieve risk-optimized collaborative segmentation. A multi-task learning framework with shared encoder and multiple expert decoders are employed as the network backbone. Mutual risk prompt learning module is developed to obtain the target-specific features and perform mutual risk recalibration between features of different targets, enabling a comprehensive understanding of the anatomical environment. The risk-recalibrated features are then fed into the task-specific gating network to adaptively activate the highly-correlated expert decoders, generating the final segmentation results. Extensive experiments conducted on both benign and malignant tumor datasets demonstrate the effectiveness of the proposed ROJS-Net. These results validate that ROJS-Net effectively resolves the optimization divergence, facilitating risk-controllable treatment planning in various clinical settings.