Yusuf Kenan Cetinoglu1, Ilker Ozgur Koska2, Muhsin Engin Uluc3, Mustafa Fazil Gelal3. 1. Batman Training and Research Hospital, Department of Radiology, 72070 Batman, Turkey. Electronic address: kenancetinoglu@hotmail.com. 2. Hatay Training and Research Hospital, Department of Pediatric Radiology, 31001 Hatay, Turkey. 3. Izmir Katip Celebi University Ataturk Training and Research Hospital, Department of Radiology, 35150 Izmir, Turkey.
Abstract
PURPOSE: Rapid detection and vascular territorial classification of stroke enable the determination of the most appropriate treatment. In this study, we aimed to investigate the performance of convolutional neural network (CNN) models in the detection and vascular territorial classification of stroke on diffusion-weighted images (DWI). METHODS: DWI of 421 cases (271 acute ischemic stroke patients and 150 cases without any ischemia findings on DWI) obtained between January 2017 to April 2020 were reviewed. We created two custom datasets. A stroke detection dataset was created with 1800 slices (900 S and 900 normal) consisting of 1400 for training, 200 for validation, 200 for test. A vascular territorial type dataset was created with 1717 slices (883 middle cerebral artery stroke, 416 posterior circulatory stroke, and 418 watershed stroke) consisting of 1117 slices for training, 300 for validation, 300 for test. A transfer learning approach based on MobileNetV2 and EfficientNet-B0 CNN architecture was used. The performance of the models was evaluated. RESULTS: Modified MobileNetV2 and EfficientNet-B0 models achieved 96% (κ: 0.92) and 93% (κ: 0.86) accuracy in stroke detection, respectively. In vascular territorial classification of stroke as middle cerebral artery, posterior circulation, or watershed infarction, an accuracy of 93% (κ: 0.895) was achieved with modified MobileNetV2 model and 87% (κ: 0.805) with modified EfficientNet-B0 CNN model. CONCLUSION: Transfer learning approach with custom top CNN models achieve sufficiently high performance for both the detection of ischemic stroke and the classification of its vascular territorial type on DWI.
PURPOSE: Rapid detection and vascular territorial classification of stroke enable the determination of the most appropriate treatment. In this study, we aimed to investigate the performance of convolutional neural network (CNN) models in the detection and vascular territorial classification of stroke on diffusion-weighted images (DWI). METHODS: DWI of 421 cases (271 acute ischemic stroke patients and 150 cases without any ischemia findings on DWI) obtained between January 2017 to April 2020 were reviewed. We created two custom datasets. A stroke detection dataset was created with 1800 slices (900 S and 900 normal) consisting of 1400 for training, 200 for validation, 200 for test. A vascular territorial type dataset was created with 1717 slices (883 middle cerebral artery stroke, 416 posterior circulatory stroke, and 418 watershed stroke) consisting of 1117 slices for training, 300 for validation, 300 for test. A transfer learning approach based on MobileNetV2 and EfficientNet-B0 CNN architecture was used. The performance of the models was evaluated. RESULTS: Modified MobileNetV2 and EfficientNet-B0 models achieved 96% (κ: 0.92) and 93% (κ: 0.86) accuracy in stroke detection, respectively. In vascular territorial classification of stroke as middle cerebral artery, posterior circulation, or watershed infarction, an accuracy of 93% (κ: 0.895) was achieved with modified MobileNetV2 model and 87% (κ: 0.805) with modified EfficientNet-B0 CNN model. CONCLUSION: Transfer learning approach with custom top CNN models achieve sufficiently high performance for both the detection of ischemic stroke and the classification of its vascular territorial type on DWI.