RADT-03. ROBUST DEEP LEARNING-BASED AUTO-SEGMENTATION FOR WHOLE VENTRICULAR AND CRANIOSPINAL IRRADIATION USING LIMITED DATA

Abstract Radiation therapy (RT) for intracranial germ cell tumor and CNS embryonal tumors warrant whole ventricular (WV) and craniospinal RT (CSI), which demands precise and complex target volume delineation for treatment planning. Recent advancements in deep learning offer the potential to improve the accuracy of segmentation processes in RT with limited number of datasets. This work employed small, limited datasets, comprising 24 cases of CT images for CSI and 25 cases of CT and MR images for WV target volume auto-segmentation, which were split into 15, 3, and 6 cases and 15, 3 and 7 cases for training, validation, and testing purposes, respectively. These datasets were trained using a widely adopted U-Net-based architecture, known as no-new-net (nnU-Net). Key functions of nnU-Net, such as network hyper-parameter optimization and specialized augmentation, possibly helped overcome challenges posed by small datasets, while enhancing segmentation accuracy. Training task was conducted within a uni-modal framework for CSI and both uni- and multi-modal frameworks for WV. The trained networks were evaluated using the dice similarity coefficient (DSC) and Hausdorff distance (HD) for 6 and 7 cases of CSI and WV target volumes, respectively. The proposed framework achieved DSCs of 98.62 for CSI target volume using CT images and 82.68, 87.54, and 87.12 for WV target volumes with CT, MR, and CT/MR combined images. For WV target volume segmentation, uni-modal training resulted in HDs of 13.28 and 15.60 for CT and MR images only, respectively. However, combining both MR and CT images improved the HD to 10.55. This work demonstrated that nnU-Net with its streamlined functions successfully trained the network to generate intracranial target volumes for RT, even with very limited data samples.