Semi-Supervised 3D Abdominal Multi-Organ Segmentation Via Deep Multi-Planar Co-Training

Abstract

In multi-organ segmentation of abdominal CT scans, most existing fully supervised deep learning algorithms require lots of voxel-wise annotations, which are usually difficult, expensive, and slow to obtain. In comparison, massive unlabeled 3D CT volumes are usually easily accessible. Current mainstream works to address semi-supervised biomedical image segmentation problem are mostly graphbased. By contrast, deep network based semi-supervised learning methods have not drawn much attention in this field. In this work, we propose Deep Multi-Planar Co-Training (DMPCT), whose contributions can be divided into two folds: 1) The deep model is learned in a co-training style which can mine consensus information from multiple planes like the sagittal, coronal, and axial planes; 2) Multiplanar fusion is applied to generate more reliable pseudolabels, which alleviates the errors occurring in the pseudolabels and thus can help to train better segmentation networks. Experiments are done on our newly collected large dataset with 100 unlabeled cases as well as 210 labeled cases where 16 anatomical structures are manually annotated by four radiologists and confirmed by a senior expert. The results suggest that DMPCT significantly outperforms the fully supervised method by more than 4% especially when only a small set of annotations is used.

Embracing imperfect datasets: A review of deep learning solutions for medical image segmentation

“…Initial annotations by Pseudo masks generated by Label noise handled by Zhang et al (2018a) K Zhou et al (2018a) propose an iterative self-learning framework, but, at each iteration, the authors train three segmentation models for the axial, sagittal, and coronal planes. Once trained, the three models scan each unlabeled 3D image slice-by-slice, generating three segmentation volumes, which are further combined through a majority voting scheme to form the final segmentation mask.…”

Multi-task Attention-Based Semi-supervised Learning for Medical Image Segmentation

“…Another semi-supervised approach is creating new pseudo labels for the unlabeled training data, such as self-training [13] and co-training [12,15], to enlist more available training resources. However, the created pseudo labels usually do not have the same quality as the ground truth for the target segmentation objective, which limits their potential for improvements from unlabeled data.…”

“…Although attention is very often applied to supervised learning (e.g., [10]), to our best knowledge, it has never been combined with semi-supervised learning. Our method has some similarities with self-training [13] and co-training [12,15], which also create new labels for the unlabeled training data on-the-fly. In contrast to these methods, our method creates labels for the reconstruction task.…”

Curriculum Semi-supervised Segmentation

“…In this paper, we focus on semi-supervised learning, a common scenario in medical imaging, where a small set of images are assumed to be fully annotated, but an abundance of unlabeled images is available. Recent progress of these techniques in medical image segmentation has been bolstered by deep learning [1,2,6,14,19,24]. Self-training is a common semi-supervised learning strategy, which consists of employing reliable predictions generated by a deep learning architecture to re-train it, thereby augmenting the training set with these predictions as pseudo-labels [1,17,18].…”