Multi-Task Learning for Integrated Automated Contouring and Voxel-Based Dose Prediction in Radiotherapy
Presenter:
Sangwook Kim, PhD Student
University of Toronto
Abstract:
Deep learning-based automated treatment planning cansignificantly enhance the efficiency and accuracy of radiotherapy. However, current planning approaches often depend on manually generated contours, limiting their efficiency. To address this, Sangwook and his colleagues implemented a multi-task learning framework that integrates automated contouring with voxel-based dose prediction, reducing the need for manual input and streamlining the planning process. Using two datasets – an in-house prostate cancer dataset and the publicly available OpenKBP head and neck cancer dataset – they developed a system that performs simultaneous segmentation and dose prediction.
Compared to conventional methods, the new framework improved the average absolute difference in dose volume histogram metrics by 2.90% for prostate cancer and 13.12% for head and neck cancer. Additionally, it enhanced dose prediction performance while maintaining high segmentation accuracy, with dice score coefficients of 0.824 for prostate and 0.716 for head and neck, compared to baseline scores of 0.818 and 0.674, respectively. These improvements can lead to more precise treatment plans and better patient outcomes.
The multi-task learning framework is not only generalizable to other anatomical sites and conditions but also holds promise for significantly reducing clinical workload and enhancing radiotherapy efficiency. By integrating automated contouring and dose prediction, this newapproach minimizes the need for sequential steps in the planning process, potentially allowing clinics to handle higher patient volumes with greater consistency and accuracy. This work illustrates the potential of AI to enable fully automated, efficient radiotherapy planning, supporting broader adoption of AI-driven tools in clinical practice.