Contour-Aware contrastive learning for 3D knee segmentation from MR images

• Published in: Pattern analysis and applications : PAA Vol. 28; no. 3
• Main Authors: Dong, Xianda, Zhang, Lei, Zhao, Xing, Zhou, Shoujun, Wang, Yuanquan, Xia, Jun, Zhang, Tao
• Format: Journal Article
• Language: English
• Published: London Springer London 01.09.2025; Springer Nature B.V
• Subjects: Attention; Cartilage; Computer Science; Contours; Decoding; Deep learning; Encoders-Decoders; Image segmentation; Knee; Modules; Original Article; Pattern Recognition; Semantics; Contour-based self-attention; Knee MRI; Segmentation; Semi-supervised learning; Contrastive learning
• ISSN: 1433-7541; 1433-755X
• DOI: 10.1007/s10044-025-01494-x

Automatic segmentation of knee MR images plays an important role in the diagnosis and treatment of knee osteoarthritis. Existing deep learning-based methods usually require considerable annotated samples, and manual labeling of knee MR images is tedious and time-consuming. To address the above problem, we propose a novel semi-supervised method, named as Contour-Aware Contrastive Learning Network (CACL-Net), for segmentation of femoral cartilage, tibial internal cartilage, and tibial external cartilage in knee MR images. The CACL-Net takes an encoder-decoder structure similar to the V-Net as backbone, and adopts a novel contrastive learning auxiliary task called Progressive Encoding Module, which can maintain the key high-level semantic information of the image and attenuate the irrelevant low-level semantic information. We also coin a contour-based self-attention module to prompt the network to pay more attention to edge details during the decoding process, thereby obtaining accurate segmentation results. Extensive experimental results demonstrate the proposed CACL-Net outperforms some other semi-supervised methods for knee MR image segmentation, and shows the potential usage of CACL-Net in the domain of semi-supervised segmentation problems. Our code is available at https://github.com/ldcdm/CLASS-Net .