Detection and vascular territorial classification of stroke on diffusion-weighted MRI by deep learning

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Rapid stroke detection and classification are crucial for effective treatment.
• Diffusion-weighted imaging (DWI) is a key modality for visualizing acute ischemic changes.
• Convolutional Neural Networks (CNNs) show promise in medical image analysis.

Purpose of the Study:

• To evaluate the performance of CNN models for stroke detection on DWI.
• To assess the capability of CNN models in classifying the vascular territory of stroke.
• To investigate the application of transfer learning using MobileNetV2 and EfficientNet-B0 architectures.

Main Methods:

• Utilized a dataset of 421 DWI cases (271 stroke, 150 normal).
• Developed custom datasets for stroke detection and vascular territorial classification.
• Employed a transfer learning approach with modified MobileNetV2 and EfficientNet-B0 CNN models.

Main Results:

• Modified MobileNetV2 achieved 96% accuracy in stroke detection (κ: 0.92).
• Modified MobileNetV2 reached 93% accuracy in vascular territorial classification (κ: 0.895).
• EfficientNet-B0 also demonstrated high performance in both detection and classification tasks.

Conclusions:

• Transfer learning with custom CNN models offers high performance for stroke detection on DWI.
• These CNN models can effectively classify the vascular territory of ischemic stroke.
• The findings support the clinical utility of AI in improving stroke diagnosis and management.