[Deep Learning-Based Automated Segmentation Algorithms of Brain and Vertebral Substructures for Radiotherapy in Pediatric Medulloblastoma].

Abstract

To evaluate the application feasibility of nnU-Net and FuseNet for automatic segmentation of pediatric medulloblastoma substructures, 60 pediatric patients who received radiotherapy and grouped by 5-year age (≤5 and >5 years) were retrospectively analyzed. Brain substructures were delineated on CT-MRI fusion images, and vertebral substructures were delineated on CT images. The three convolutional neural network models of U-Net, nnU-Net and FuseNet were trained (24 cases/group) and tested/validated (6 cases/group), with 20 external cases verifying generalization. The three models and Atlas-based methods were compared using DSC; nnU-Net/FuseNet's HD95, RAVD and manual correction time were evaluated. Results showed that FuseNet performed best in brain segmentation, outperforming Atlas and U-Net in vertebral substructures for both age groups ( P=0.028/0.005 and 0.005/0.005) but not differing from nnU-Net ( P=0.107/0.236). Its DSC exceeded 0.8 for most substructures (except cerebellar anterior lobe/hippocampus in group≤5 years; hippocampus in group>5 years), with shortest correction time in both age groups. This study has demonstrated that nnU-Net can achieve good segmentation; FuseNet can improve brain segmentation accuracy via dynamic multimodal feature fusion, with the highest correction efficiency.