AbstractINTRODUCTIONSegmenting brain structures around a tumor on brain images is important for radiotherapy and surgical planning. Current auto-segmentation methods often fail to segment brain anatomy when it is distorted by tumors.OBJECTIVETo develop and validate 3D capsule networks (CapsNets) that can segment brain structures with novel spatial features that were not represented in the training data.MethodsWe developed, trained, and tested 3D CapsNets using 3430 brain MRIs acquired in a

more »

... -institutional study. We compared our CapsNets with U-Nets using multiple performance measures, including accuracy in segmenting various brain structures, segmenting brain structures with spatial features not represented in the training data, performance when the models are trained using limited data, memory requirements, and computation times.RESULTS3D CapsNets can segment third ventricle, thalamus, and hippocampus with Dice scores of 94%, 94%, and 91%, respectively. 3D CapsNets outperform 3D U-Nets in segmenting brain structures that were not represented in the training data, with Dice scores more than 30% higher. 3D CapsNets are also remarkably smaller models compared to 3D U-Nets, with 93% fewer trainable parameters. This led to faster convergence of 3D CapsNets during training, making them faster to train compared to U-Nets. The two models were equally fast during testing.CONCLUSION3D CapsNets can segment brain structures with high accuracy, outperform U-Nets in segmenting brain structures with features that were not represented during training, and are remarkably more efficient compared to U-Nets, achieving similar results while their size is one order of magnitude smaller.