A Deep Learning System for Diagnosing Ischemic Stroke by Applying Adaptive Transfer Learning

Abstract

A stroke is the most common, very dangerous single-organ disease and aggravates social burden in the aging society. The stroke can be tested through a variety of imaging methods, among which a test method using CT imaging is known to deal promptly with an emergency patient in the early stage of stroke. Diagnosing ischemic stroke using CT images has advantages such as fewer spatial constrains and quick shooting time. However, diagnosis through images is very difficult, which is a major disadvantage of this method. This study proposed a deep learning system that can conduct learning and classification for ischemic stroke, which is a small dataset and hard to conduct image data learning. This study also proposed a pre-processing algorithm optimized for ischemic stroke based on the non-contrast CT data from the middle cerebral artery (MCA) region. Additionally, this study suggested adopting the adaptive transfer learning algorithm that optimizes the transfer learning module to overcome the problem of insufficient data while training neural networks. When stroke was diagnosed using the proposed system, the performance of it was 18.72% better than the existing system.