Enhancing Ischemic Brain Stroke Detection on CT Images: A Investigation of Transfer Learning Techniques of DenseNet-201 for Neuroimaging Analysis

Over the past few decades, strokes, also called cerebrovascular diseases or cerebral vascular accidents (CVAs), have become a significant contributor to morbidity and mortality globally. Brain stroke early diagnosis is vital to initiate treatment to prevent brain cells from becoming necrotic. Non-contrast-enhanced computed tomography (NCCT) is the primary imaging technique used to diagnose brain stroke conditions. Deep learning-based computer-aided diagnoses are developed to aid in diagnosis as ‘the brain is time’. DenseNet201 architecture has alleviated the issue of gradient vanishing by utilizing dense connections. This study aims to detect brain stroke conditions by developing two deep learning models utilizing the transfer learning architecture of DenseNet201. Two deep learning models serve as a two-step process to identify ischemic stroke and its subtype conditions. Model 01 achieved an intrinsic accuracy of 97. 03% in the training set, 94. 50% in the validation set. Also, test and test time augmentation accuracy (TTA) was 94.50%. Model 02 reached an intrinsic accuracy of 98.02% in the training set and 95.67% in the validation set. Further, the test and TTA accuracy were 95.67%.