A New Deep-Net Architecture for Ischemic Stroke Lesion Segmentation

Abstract

Ischemic stroke, brain cells death due to a lack of oxygen, is a leading cause of long-term disability and death. Accurate diagnosis and timely intervention can effectively improve the blood supply of the ischemic stroke area and minimize brain damage. Recent studies have shown the potential to use magnetic resonance imaging (MRI) to provide contrast imaging to visualize and detect lesions. However, manual segmentation of the stroke lesion produced by MRI is a tedious and time-consuming task. Therefore, the automatic ischemic stroke lesion segmentation method may show excellent advantages. In this paper, we propose a novel deep learning method used to detect and localize brain ischemic stroke, a generalization encoderdecoder by modifying U-Net architecture. We integrate multi-path architecture into both encoder and decoder blocks to captures different levels of the encoded state, which helps in more robust decision-making for stroke lesion segmentation. In bottleneck of the architecture, we applied dilated blocks to improve the underlying predictive capabilities. The proposed method has been tested on the publicly accessible web platform provided by the MICCAI Ischemic Stroke Lesion Segmentation (ISLES) challenge. The results demonstrate that the proposed method achieves a mean dice coefficient 0.91 of with the training and 0.84 with the testing data respectively.