Semi-Supervised Knee Cartilage Segmentation With Successive Eigen Noise-Assisted Mean Teacher Knowledge Distillation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Knee Osteoarthritis (OA) diagnosis relies on accurate knee cartilage segmentation from MRI.
• Domain shifts from varying MRI scanners pose significant challenges for cross-modality segmentation.
• Current methods struggle with feature discrimination and capturing higher-order correlations.

Purpose of the Study:

• To develop a novel framework for robust 2D knee MRI cross-modality adaptation.
• To improve knee cartilage segmentation accuracy for Osteoarthritis diagnosis.
• To address limitations of existing methods in handling domain shifts and subtle features.

Main Methods:

• Proposed Successive Eigen Noise-assisted Mean Teacher Knowledge Distillation (SEN-MTKD) framework.
• Utilized Eigen Low-rank Subspace (ELRS) for pseudo-label generation and domain-invariant features.
• Incorporated Successive Eigen Noise (SEN) for enhanced discrimination and data perturbation.
• Implemented a subspace-based feature distillation loss (LRBD) for robust representation.

Main Results:

• SEN-MTKD demonstrated superior performance over state-of-the-art benchmarks on public and private datasets.
• The framework effectively handles domain shifts and improves segmentation of less prominent cartilages.
• Achieved robust feature representation and reliable pseudo-labeling through advanced distillation.

Conclusions:

• SEN-MTKD offers a significant advancement in cross-modality knee MRI segmentation for OA diagnosis.
• The proposed method enhances discrimination and captures critical semantic information across domains.
• This framework provides a reliable solution for adapting MRI data from diverse scanning technologies.