MTKD-LRS: Semi Supervised Knee Cartilage Segmentation Using Eigen Low Rank Subspace Assisted Mean-Teacher Framework

Segmentation of cartilages to examine the Knee osteoarthritis is a challenging problem in medical image analysis, due to distribution gaps in source and target domains. Existing models trained on one MRI domain struggle to generalize to MRI scans produced from different scanners, highlighting the need to develop novel approaches to adapt to cross-modalities. In this paper, we propose a novel Eigen Low-rank subspace-assisted Mean Teacher Knowledge Distillation framework (MTKD-LRS) using a semi-supervised learning approach. This framework leverages the low-rank approximations within the deep feature subspaces to capture meaningfullatent patterns and construct the domain-invariant feature representations. By preserving robust feature maps associated with larger singular values and leveraging the lower singular value feature maps as successive truncated noise, the student model is optimized with more robust supervision to bridge the gap between the cross-modality MRI data. Extensive experiments on public and private datasets demonstrate the effectiveness of MTKD-LRS over existing state-of-the-art approaches.