Deep Learning Technique for Detecting and Analysing Ischemic Stroke Using MRI Images

Abstract

The quantitative analysis of cerebral MRI images plays a pivotal role in stroke diagnosis and treatment. Deep learning, particularly CNNs, with their robust learning capabilities, offer an effective tool for lesion detection. To address the unique properties of stroke injuries and automate detection processes, we compiled a dataset of brain MRI images from various medical sources, representing patients affected by ischemic strokes. Different deep learning-based networks, including “Single Shot Multibox Detector (SSD)”, “Region-based CNN with ResNet101 (RCNN-ResNet101)”, “RCNN with VGG16 (RCNN- VGG16)”, and “YOLOV3”, were employed for automated lesion detection. The evaluation focused on achieving optimal precision in comparison to existing methods across Diffused Weight, Flair, and T1 modalities of MRI datasets. The developed technique involves extracting deep features during the encoding stage, followed by the minimization of features using fully connected layers. Significant handcrafted features, such as Local Binary Pattern (LBP) and Gray Level Co-occurrence Matrix (GLCM), were incorporated alongside deep features. The concatenation of these features was implemented to maximize the dimension of the feature vector. This concatenated vector was then used to train and test the performance of various classifiers. Binary classification was employed to categorize brain images into normal or stroke affected. Initially, SoftMax was used as the default classifier. The performance of each classifier was individually evaluated, and the best-performing classifier was selected to confirm the overall effectiveness of the proposed technique. This all-encompassing strategy not only leverages deep learning for automatic lesion detection but also integrates handcrafted features and diverse classifiers to improve the precision and dependability of stroke detection across various brain MRI image modalities.