End to end vision transformer architecture for brain stroke assessment based on multi-slice classification and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Brain stroke is a major cause of global disability and mortality.
• Current diagnosis relies on subjective radiologist interpretation of CT scans, leading to potential errors.
• Accurate and automated methods for stroke diagnosis and localization are crucial for improving patient outcomes.

Purpose of the Study:

• To enhance the Vision Transformer (ViT) architecture for multi-slice classification of brain CT scans.
• To develop an automated system for classifying CT scans into Normal, Infarction, and Hemorrhage categories.
• To achieve patient-wise stroke localization using an end-to-end ViT framework.

Main Methods:

• Modified the Vision Transformer (ViT) with neural network layers for multi-slice CT scan classification.
• Utilized ViT and convolutional neural network layers for stroke detection and bounding box localization.
• Employed a patient-wise, multi-slice approach for comprehensive stroke analysis.

Main Results:

• Achieved an overall accuracy of 87.51% in classifying brain CT scan slices.
• Demonstrated high precision in patient-wise stroke localization.
• The framework shows potential for accurate and reliable stroke diagnosis and localization.

Conclusions:

• Enhanced ViT architecture offers automated stroke diagnosis and localization from CT scans.
• Deep learning provides an objective and consistent approach, potentially enabling personalized treatment.
• Further validation on diverse datasets is needed to confirm clinical utility.