Deep learning-based multi-organ segmentation for pediatric radiotherapy

Abstract

Organ delineation is an essential but time-consuming step in pediatric kidney tumor radiotherapy. Interest in the automation of this task is increasing, and this work attempts to automatize abdominal multi-organ delineation for external beam radiotherapy. Convolutional neural networks, specifically semantic image segmentation, will be used. Different architectures will be explored, considering U-Net and nnU-Net as baseline approaches, and comparing their performances to some of the latest proposed networks, i.e., adopting attention-gating and vision-transformers. The top-performing methods obtain a median DSC above 0.95 for all organs except for the pancreas (with a DSC of 0.85). The effectiveness of attention gating is demonstrated by enhancing the performance of the baseline models. Additionally, the top-performing model achieves DSCs above 0.90 across the full range of patient-ages and organ volumes (except for the pancreas and the bowel).