Multi-class segmentation of knee MRI based on hybrid attention

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate knee MRI segmentation is vital for diagnosing knee conditions.
• Existing methods struggle with class imbalance and small structure feature detection.

Purpose of the Study:

• To improve multi-class segmentation of knee MRI images.
• To address limitations of existing segmentation methods, particularly class imbalance and feature extraction for small structures.

Main Methods:

• Developed a Hierarchical Feature Enhancement Fusion (HFEF) module for multi-level feature integration.
• Introduced an Atrous Squeeze Attention (ASA) module for multi-scale feature focus and long-range dependency capture.
• Optimized the loss function to handle class imbalance and limited data.

Main Results:

• The HASA-ResUNet model achieved a 12.12% improvement in Intersection over Union (IoU) for the anterior cruciate ligament.
• Demonstrated a 3.32% improvement in mean Intersection over Union (mIoU) across all classes compared to U-Net.
• Showcased enhanced segmentation performance for low-frequency and small-sized classes.

Conclusions:

• The proposed hybrid attention and multi-scale strategy effectively addresses class imbalance in knee MRI segmentation.
• The HASA-ResUNet model significantly improves segmentation accuracy for critical knee structures.
• This approach enhances overall segmentation performance for degenerative knee disease and sports injury analysis.