Semi-supervised multi-organ segmentation via multi-planar co-training.

Multi-organ segmentation is a critical problem in medical image analysis due to its great value for computer-aided diagnosis, computer-aided surgery, and radiation therapy. Although fully-supervised segmentation methods can achieve good performance, they usually require a large amount of 3D data, such as CT scans, with voxel-wised annotations which are usually difficult, expensive, and slow to obtain. By contrast, large unannotated datasets of CT images are available. Inspired by the well-known semi-supervised learning framework co-training, we propose multi-planar co-training (MPCT), to generate more reliable pseudo-labels by enforcing consistency among multiple planes, i.e., saggital, coronal, and axial planes, of 3D unlabeled medical data, which play a vital role in our framework. Empirical results show that generating pseudo-labels by the multi-planar fusion rather than a single plane leads to a significant performance gain. We evaluate our approach on a new collected dataset and show that MPCT boosts the performance of a typical segmentation model, fully convolutional networks, by a large margin, when only a small set of labeled 3D data is available, i.e., 77.49% vs. 73.14%.