CAN3D: Fast 3D medical image segmentation via compact context aggregation

Wei Dai, Boyeong Woo, Siyu Liu, Matthew Marques, Craig Engstrom, Peter B. Greer, Stuart Crozier, Jason A. Dowling, Shekhar S. Chandra

Abstract: Highlights•We propose a neural network with a highly compact structure for full-size 3D MR image semantic segmentation.•We propose a loss function to restrain the absolute values during the training, which improves the targets’ volumetric and surface accuracy.•Our model achieves state-of-the-art performance with faster computation time, fewer parameters, and a smaller memory footprint than other CNN models.•We released our expert manual semantic labelled 3D image data for the pelvis organs and prostate consisting of 211 images.