Purpose: There have been suggestions about different ischemic lesion patterns in different stroke mechanisms. Recent advances in artificial intelligence techniques have led to the application of deep learning algorithm to the analysis of neuroimaging. The present study sought to classify stroke subtypes based on the ischemic lesion patterns on magnetic resonance imaging using pattern recognition techniques of deep learning algorithm.
Methods: We tried a ternary classification using 3 dimentional convolutional neural networks to classify the diffusion-weighted magnetic resonance imaging scans from patients with presumed large artery atherosclerosis (LAA), cardioembolism (CE), and small vessel occlusion (SVO) subtypes. The dataset consisted of 1,200 diffusion-weighted magnetic resonance imaging cases. Images were normalized to the MNI 152 template. Without any information about the clinical variables and vascular imaging, we trained, validated and tested our 3 dimensional convolutional networks algorithm. To reduce overfitting, we applied dropout regulation and batch normalization.
Results: We used 720 cases for training and 120 cases for validation. For CE vs. SVO, shown on a test set of 240 cases, our 3 dimensional convolutional networks algorithm correctly classified the test set into their respective groups with a mean accuracy of 78.9% and a standard deviation of 3.1%. For CE vs. LAA, shown on a test set of 240 cases, the algorithm correctly classified the test set into their respective groups with a mean accuracy of 65.4% and a standard deviation of 2.8%. For LAA vs. SVO, shown on a test set of 240 cases, the algorithm correctly classified the test set into their respective groups with a mean accuracy of 65.6% and a standard deviation of 2.9%. For ternary classification of CE, LAA, and SVO, shown on a test set of 360 cases, the algorithm correctly classified the test set with a mean accuracy of 54.0% and a standard deviation of 4.4%.
Conclusion: The results show the feasibility of stroke subtype classification based on ischemic lesion patterns analysis and indicate a role for 3 dimensional convolutional neural networks algorithm in acute ischemic stroke imaging.

Lesion patterns recognition using 3 dimensional convolutional neural networks for automated classification of stroke subtypes