A feature-enhanced network for stroke lesion segmentation from brain MRI images

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Neurology

Background:

• Accurate segmentation of stroke lesions is crucial for timely medical diagnosis and treatment.
• Current deep learning methods face limitations in utilizing local features and preserving semantic information during segmentation.

Purpose of the Study:

• To propose a novel Feature Refinement and Protection Network (FRPNet) for accurate and efficient stroke lesion segmentation.
• To address the limitations of existing deep learning models in capturing both global and local features and preventing information loss.

Main Methods:

• Developed FRPNet with a symmetric encoding-decoding structure.
• Incorporated Twin Attention Gate (TAG) module for global and local feature extraction using self-attention and bi-directional attention.
• Integrated Multi-dimension Attention Pooling (MAP) module to mitigate feature loss during encoding.

Main Results:

• FRPNet significantly outperformed state-of-the-art methods on two ischemic stroke datasets.
• Achieved 60.16% DSC and 36.20px HD on one dataset, and 85.72% DSC and 27.02px HD on another.
• Demonstrated excellent efficacy and generalizability across various stroke stages and image sequences.

Conclusions:

• The proposed FRPNet effectively enhances stroke lesion segmentation accuracy and efficiency.
• The novel TAG and MAP modules successfully address limitations in feature extraction and information preservation.
• FRPNet shows strong potential for clinical application in stroke diagnosis and treatment planning.