BRIEFnet: Deep Pancreas Segmentation Using Binary Sparse Convolutions

Dense prediction using deep convolutional neural networks (CNNs) has recently advanced the eld of segmentation in computer vision and medical imaging. In contrast to patch-based classication, it requires only a single path through a deep network to segment every voxel in an image. However, it is dicult to incorporate contextual information without using contracting (pooling) layers, which would reduce the spatial accuracy for thinner structures. Consequently, huge receptive elds are required which might lead to disproportionate computational demand. Here, we propose to use binary sparse convolutions in the rst layer as a particularly eective approach to reduce complexity while achieving high accuracy. The concept is inspired by the successful BRIEF descriptors and complemented with 11 convolutions (cf. network in network) to further reduce the number of trainable parameters. Sparsity is in particular important for small datasets often found in medical imaging. Our experimental validation demonstrates accuracies for pancreas segmentation in CT that are comparable with state-of-the-art deep learning approaches and registration based multi-atlas segmentation with label fusion. The whole network, which also includes a classic CNN path to improve local details, can be trained in 10 minutes. Segmenting a new scan takes 3 seconds even without using a GPU.