MambaMatch: A Novel Model for Semi-Supervised Spinal Cortical and Cancellous Bone Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Precise segmentation of cortical and cancellous bone is crucial for safe laminectomy procedures.
• Limited annotations and high anatomical similarity present significant challenges in current bone segmentation methods.

Purpose of the Study:

• To propose MambaMatch, an end-to-end semi-supervised segmentation framework to address challenges in bone segmentation.
• To enhance feature extraction and diversity for improved segmentation accuracy.

Main Methods:

• Utilized a teacher-student paradigm with a Mamba-ASPP-Unet (MAU) student module integrating multi-scale Atrous Spatial Pyramid Pooling (ASPP) and attention mechanisms.
• Implemented a dual-stream perturbation strategy including Correlation-Guided CutMix Augmentation (CGCA) and standard augmentations.
• Introduced dynamic thresholding and temperature scaling to adaptively manage pseudo-label reliability and consistency loss.

Main Results:

• Achieved 79.69% mIoU and 83.57% Dice on the CT Cortical-Cancellous Dataset.
• Demonstrated strong robustness across diverse datasets including MRI-SPIDER, SKIN-ISIC 2018, PH2, and CT-VerSe.
• The MambaMatch model proved to be lightweight and efficient.

Conclusions:

• MambaMatch offers an efficient and accurate solution for bone segmentation, with potential to support clinical workflows in spinal surgery.
• The framework shows broad applicability across various clinical imaging scenarios.