Ischemic stroke lesion segmentation using stacked sparse autoencoder

Abstract

Automatic segmentation of ischemic stroke lesion volumes from multi-spectral Magnetic Resonance Imaging (MRI) sequences plays a vital role in quantifying and locating the lesion region. Most existing methods mainly rely on designing hand-crafted features followed by a classifier model for ischemic stroke lesion segmentation. Design of these features requires complex domain knowledge and often lacks the ability to differentiate between the stroke lesions and the normal classes. In this work, we propose an unsupervised featured learning approach based on stacked sparse autoencoder (SSAE) framework for automatically learning the features for accurate segmentation of stroke lesions from brain MR images. A deep architecture is designed using sparse autoencoder (SAE) layers, followed by support vector machine (SVM) classifier for classifying the patches into normal or lesions. We validated our approach on a publicly available Ischemic Stroke Lesion Segmentation (ISLES) 2015 dataset, with a mean precision of 0.968, mean dice coefficient (DC) of 0.943, mean recall of 0.924 and mean accuracy of 0.904. The experimental results show that our proposed approach significantly outperforms the state-of-the-art methods in terms of precision, DC, and recall. Quantitative evaluation was carried out and compared with the existing approaches, which demonstrates that the proposed method is 25.71%, 36.67%, and 16.96% higher in terms of precision, DC and recall values, respectively. The unsupervised features learned via SSAE framework performs better than the hand-crafted features and can be easily trained on large datasets.