Author: Alexander Choi, William Ross Green, Christine Hill-Kayser, Gary D. Kao, Michael LaRiviere, Rafe A. McBeth, Steven Philbrook 👨🔬
Affiliation: Department of Radiation Oncology, University of Pennsylvania 🌍
Purpose: To validate the potential of clinical deployment of an in-house AI-driven auto-segmentation tool for pediatric craniospinal irradiation (CSI) in proton therapy, with goals of reducing manual contouring time, improving workflow efficiency, and ensuring consistent delineation accuracy in a high-volume clinical setting.
Methods: Segmentation data from 19 pediatric proton CSI patients were retrospectively collected to train and evaluate a custom model using the nnU-Net framework. The model was trained to segment target structures, including the thecal sac (with nerve roots) and the brain, following clinical guidelines. To assess its performance, the model was tested on data from five patients outside the training set, with accuracy evaluated using the Dice similarity coefficient. For two test cases, the time required for physicians to review and refine the AI-generated contours was recorded and compared to manual contouring time. Workflow efficiency, including overall time savings, was quantified to assess the model’s potential to streamline clinical operations.
Results: The AI-driven segmentation achieved mean Dice coefficients of 0.97 for the thecal sac and 0.99 for the brain. Physician review and correction of AI-generated contours averaged 5 minutes per patient, compared to 23 minutes for manual contouring using commercial auto-segmentation software as a starting point.
Conclusion: This study demonstrates the feasibility and effectiveness of an in-house AI-based segmentation tool for pediatric craniospinal irradiation. It significantly reduces contouring time while maintaining high segmentation accuracy, streamlining the process and enabling faster treatment planning. These results highlight its potential to improve workflow efficiency in high-volume clinical settings and contribute to shorter patient turn-around times. This work underscores the value of developing AI-driven models tailored to clinical needs, paving the way for broader integration of AI into radiotherapy to enhance efficiency, precision, and quality of care.