Brain stroke detection and classification using deep learning-based transCBAMSegNet and hybrid transformer models

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Stroke necessitates rapid assessment to minimize mortality and disability.
• Accurate brain lesion segmentation and stroke type classification from CT scans are crucial for timely clinical decisions.

Purpose of the Study:

• To develop a deep learning framework for automated stroke lesion segmentation and CT-based stroke type classification.
• To create a computer-aided tool for early-stage, imaging-based stroke assessment.

Main Methods:

• A deep learning framework incorporating TransCBAMSegNet for lesion segmentation.
• A hybrid MaxxViT-ViT-SwinV2 feature fusion strategy for stroke type classification.
• Utilized computed tomography (CT) imaging data for analysis.

Main Results:

• The TransCBAMSegNet model achieved 99.60% accuracy, 75.98% mIoU, and 75.50% DSS for segmentation.
• The classification stage reached 97% accuracy in differentiating ischaemic stroke, haemorrhagic stroke, and normal cases.
• The multi-transformer fusion approach outperformed single-backbone configurations.

Conclusions:

• The proposed framework provides a robust and reproducible method for automated stroke analysis using CT imaging.
• The developed tool can support early-stage clinical decision-making in stroke management.
• Deep learning models demonstrate significant potential in enhancing stroke assessment accuracy and efficiency.