Towards Clinical Diagnosis: Automated Stroke Lesion Segmentation on Multi-Spectral MR Image Using Convolutional Neural Network

Abstract

The patient with ischemic stroke can benefit most from the earliest possible definitive diagnosis. While a quantitative evaluation of the stroke lesions on the magnetic resonance images (MRIs) is effective in clinical diagnosis, manually segmenting the stroke lesions is commonly used, which is, however, a tedious and time-consuming task. Therefore, how to segment the stroke lesions in a fully automated manner has recently extracted extensive attentions. Considering that the clinically acquired MRIs usually have thick slices, we propose a 2D-slice-based segmentation method. In particular, we use multi-spectral MRIs, i.e., diffusion weighted image, apparent diffusion coefficient, and T2-weighted image, as input, and propose a residual-structured fully convolutional network (Res-FCN). The proposed Res-FCN is trained and evaluated on a large data set with 212 clinically acquired MRIs, which achieves a mean dice coefficient of 0.645 with a mean number of false negative lesions of 1.515 per subject. The proposed Res-FCN is further evaluated on a public data set, i.e., ISLES2015-SISS, which presents a very competitive result among all 2D-slice-based segmentation methods.

Keywords

• Deep learning
• stroke lesion segmentation
• residual network (ResNet)
• convolutional neural network (CNN)
• fully convolutional network (FCN)