Multi-resolution Path CNN with Deep Supervision for Intervertebral Disc Localization and Segmentation

Automatic localization and segmentation of intervertebral discs (IVDs) from MR images plays a vital role in the diagnosis of pathological changes of IVDs. In this paper, we present a novel multi-resolution path network with deep supervision (MRP-DSN) to handle this challenging task. The MRP-DSN is based on a multi-scale backbone network, which is a DenseNet with densely connected atrous spatial pyramid pooling. More importantly, we introduce a multi-path network architecture that treats the segmentation of IVDs as a multi-task problem, i.e. segments IVDs into label maps at multiple resolutions, and then integrates them together for predicting the overall segmentation results. Each path is independently initialized and have a specific objective under deep supervision, which makes the training of each path more effective without interfering each other, thus results in more robust segmentation. We further design a training strategy that can eliminate the influence of unlabeled thoracic discs and make the training focus on the spine area. We evaluated our method on MICCAI 2018 IVDM3Seg Challenge dataset, the proposed MRP-DSN achieves superior performance.