Shape-Guided Dual Consistency Semi-Supervised Learning Framework for 3-D Medical Image Segmentation

Abstract

Popular semi-supervised 3D medical image segmentation networks commonly suffer from two limitations: First, the geometry shape constraint of targets is frequently disregarded, leading to coarse segmentation results. Second, semi-supervision is only performed on the last layer of the decoder, resulting in the insufficient representation learning of 3D convolution neural network. To address these issues, we propose a shape-guided dual consistency semi-supervised learning (SDC-SSL) framework for 3D medical image segmentation. Indeed, the proposed framework has two dominating advantages. Initially, a geometry-aware shape constraint is presented and used to learn the shape representation, which converts the differences between two networks into an unsupervised loss and lets the framework learn the boundary distance information of targets in unlabeled challenging regions. Additionally, a deep-supervised knowledge transfer strategy is developed and employed by the proposed framework, which can upgrade the generalization ability of our framework without increasing any extra parameters and computation costs in the inference phase. Experimental results demonstrate that the proposed framework outperforms state-of-the-art methods on two challenging 3D medical image segmentation tasks due to effective geometry-aware shape constraint on unlabeled data and the strong ability of knowledge mining on labeled data. The code is available at: https://github.com/SUST-reynole/SDC-SSL.