KCB-Net: A 3D knee cartilage and bone segmentation network via sparse annotation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate knee cartilage and bone segmentation is vital for diagnosing knee osteoarthritis (OA) and articular damage.
• Deep learning (DL) methods excel at medical image segmentation but require substantial annotated data, which is costly and time-consuming for 3D images.

Purpose of the Study:

• To develop a novel knee cartilage and bone segmentation framework (KCB-Net) for 3D MR images utilizing sparse annotations.
• To bridge the performance gap between sparse and fully annotated datasets in medical image segmentation.

Main Methods:

• KCB-Net employs an unsupervised scheme to select the most representative slices for annotation.
• It utilizes an ensemble model trained on annotated slices, followed by self-training with pseudo-labels enhanced by a bi-directional hierarchical earth mover's distance (bi-HEMD) algorithm.
• Segmentation results are fine-tuned using the primal-dual Internal Point Method (IPM).

Main Results:

• KCB-Net demonstrates superior performance compared to state-of-the-art methods, even with full annotation.
• The framework achieves high-quality segmentation results with annotation ratios as low as 10% across four diverse 3D MR knee joint datasets.
• Experiments validated on SKI10, OAI ZIB, Iowa, and iMorphics datasets.

Conclusions:

• KCB-Net offers an effective solution for 3D knee MR image segmentation with significantly reduced annotation requirements.
• The proposed framework holds promise for improving the efficiency and accessibility of OA diagnosis and analysis.
• Sparse annotation strategies combined with advanced DL techniques can achieve expert-level segmentation performance.