DRRN: Differential rectification & refinement network for ischemic infarct segmentation

Abstract

Accurate segmentation of infarct tissue in ischemic stroke is essential to determine the extent of injury and assess the risk and choose optimal treatment for this life-threatening disease. With the prior knowledge that asymmetric analysis of anatomical structures can provide discriminative information, plenty of symmetry-based approaches have emerged to detect abnormalities in brain images. However, the inevitable non-pathological noise has not been fully alleviated and weakened, leading to unsatisfactory results. A novel differential rectification and refinement network (DRRN) for the automatic segmentation of ischemic strokes is proposed. Specifically, a differential feature perception encoder (DFPE) is developed to fully exploit and propagate the bilateral quasi-symmetry of healthy brains. In DFPE, an erasure-rectification (ER) module is devised to rectify pseudo-lesion features caused by non-pathological noise through utilising discriminant features within the symmetric neighbourhood of the original image. And a differential-attention (DA) mechanism is also integrated to fully perceive the differences in cross-axial features and estimate the similarity of long-range spatial context information. In addition, a crisscross differential feature reinforce module embedded with multiple boundary enhancement attention modules is designed to effectively integrate multi-scale features and refine textual details and margins of the infarct area. Experimental results on the public ATLAS and Kaggle dataset demonstrate the effectiveness of DRRN over state-of-the-arts.