Generalizability of Deep Learning Models on Brain Tumour Segmentation

Brain tumour segmentation is a hard and time-consuming task to be conducted in the process of radiotherapy planning. Deep Learning (DL) applications have a significant improvement in image segmentation tasks. In this work, we apply DL models such as 2D and 2.5D U-NET to the segmentation task of a brain tumour on the BraTS 2021 dataset and our local dataset. The 2.5D network is a modified version of 2D U-NET by using three slices as an input for each magnetic resonance imaging (MRI) sequence. We achieve the best segmentation results with 2.5D U-NET on BraTS with Dice scores of 86.97%, 91.27% and 94.42% for enhancing tumour, tumour core and whole tumour respectively. On the other hand, our best segmentation result of the GTV delineation on the local dataset is a Dice score of 78.51% for 2D U-NET. Although the result of GTV contours is not improved by 2.5D for the local dataset due to non-fixed voxel size, the Dice scores of ET, TC and WT are improved by the proposed 2.5D U-NET for the BraTS dataset.