A Multiple Encoders Network for Stroke Lesion Segmentation

Abstract

Medical image segmentation plays a vital role in clinical application of auxiliary diagnosis. Having said that, it is still a challenging task to accurately segment lesion areas from case images, because lesions are always different in size and shape, and the boundaries of lesion area is usually too rough, which greatly limits the accuracy of the current mainstream automatic segmentation algorithms. The lesions in stroke data are mostly small targets, and the boundary area are very tortuous, which greatly increases the difficulty of segmentation. Therefore, there is still a lot of space for segmentation of stroke data sets. In this paper, we propose a novel multiple encoders network for stroke lesion segmentation, called ME-Net, which is mainly composed of the atrous spatial pyramid pooling (ASPP) module and residual encoder. Particularly, ASPP uses a parallel structure of porous convolutions with different sampling rates to obtain context information for multi-scale features. The residual encoding can obtain high-level semantic features at high resolution. In addition, ME-Net adopts an encoding and decoding framework. We divided the pathological images into three levels according to the number of pixels in the masked area and evaluated our proposed method on an open dataset Anatomical Tracings of Lesions After Stroke (ATLAS) with encouraging performance achieved compared to other state-of-the-art approaches.