Learning From Multimodal Privileged Information to Predict Radiation Dose for Glioblastoma Radiotherapy

Abstract

Glioblastoma (GBM) is one of the most common and aggressive malignant brain tumors, with radiotherapy being the primary treatment modality. Designing an optimal radiotherapy plan, however, is often a labor-intensive process. To accelerate the planning process, several dose prediction methods have been developed with promising outcomes. Yet, existing dose prediction methods highly rely on the contouring of organs-at-risks (OARs) and tumors, which are typically cumbersome and time-consuming. In this study, we propose a mask-free dose prediction method leveraging multimodal privileged information within the framework of generalized knowledge distillation. The proposed network is a teacher-student model, and the anatomical masks are treated as privileged information, which is available for training the teacher model but not for the student model during inference. The pretrained teacher model transfers the learned anatomical information to the student model via knowledge distillation. To ensure the conformity of the predicted dose map, we design a high-dose contour loss. To the best of our knowledge, this is the first approach that learns from multimodal information to predict radiation dose distribution without requiring contour structures. Our experimental results demonstrate that incorporating multimodal information significantly assists the model in learning the tumor characteristics, even in the absence of anatomical contours.