Enhancing lumbar disc herniation classification through region-of-interest guidance and geometric shape features

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Spine Surgery

Background:

• Lumbar disc herniation (LDH) is a prevalent spinal degenerative disease.
• Magnetic Resonance Imaging (MRI) is crucial for LDH detection.
• Current computer-aided diagnosis (CAD) methods struggle with varied disc shapes, blurred boundaries, and unclear classification, hindering accurate lesion differentiation.

Purpose of the Study:

• To develop an enhanced computer-aided diagnosis system for lumbar disc herniation (LDH).
• To improve the classification accuracy and segmentation performance for LDH using MRI.
• To address the challenges posed by complex disc morphologies and ambiguous lesion types in current CAD systems.

Main Methods:

• Proposed RGGS-Net (Region-of-Interest Guidance and Geometric Shape features Network) for enhanced LDH classification.
• Integrated diseased lumbar disc segmentation with lesion type classification.
• Implemented a region-of-interest guided module with region-of-interest supervision for feature refinement.
• Utilized weighted skip connections for feature balancing and hierarchical supervision for deep decoder training.

Main Results:

• Achieved a classification accuracy of 0.965 for LDH.
• Reached a Dice score of 0.957 for vertebrae and disc segmentation.
• Demonstrated the effectiveness of RGGS-Net in numerous experimental evaluations.

Conclusions:

• RGGS-Net effectively enhances LDH classification by leveraging geometric features and region-of-interest guidance.
• The proposed method improves both the classification of LDH lesion types and the segmentation accuracy of spinal structures.
• RGGS-Net offers a promising solution for challenging computer-aided diagnosis of lumbar disc herniation.