DEEP LEARNING-BASED BRAIN STROKE DETECTION: CLINICAL RELIABILITY ANALYSIS OF CNN AND HYBRID NEURORESUNET ARCHITECTURES

Abstract

Brain stroke is a leading cause of death and long-term disability in the world and timely and accurate diagnosis is critical to better patient outcomes and decreasing mortality rates. The automated detection of stroke using deep learning has been a new promising way to assist clinical decision making and improve the efficiency of diagnosis. Two deep-learning models: a Convolutional Neural Network (CNN) and a Hybrid NeuroResUNet model, are compared in this study for brain stroke automated prediction by computed tomography (CT) scan images. A labeled dataset consisting of stroke and non-stroke CT images was used for training and evaluation of the models. Standard Evaluation metrics such as accuracy, precision, recall, specificity, F1 score and confusion matrix analysis were used to evaluate performance. The experimental results showed that the Hybrid NeuroResUNet outperforming the other methods with an overall accuracy of 95.06%; however, the method had a relatively high false-negative rate which could pose risks for clinicians using the system for stroke detection. The CNN model showed better accuracy of 93.94%, sensitivity of 96.77%, specificity of 91.43%, and F1 score of 93.75%, compared to the others. The CNN model's performance in accurate classification of stroke and non-stroke cases was also verified using a confusion matrix analysis. The results showed that although the CNN model had a slightly lower accuracy, it was more clinically reliable and robust in automatic brain stroke detection and thus proved to be more suitable for clinical applications.