Joint Image-Mask Conditional Diffusion for Robust Meniscus Segmentation in Knee MR Images

Successful meniscal allograft transplantation requires grafts that precisely match the native meniscus in size and shape, which necessitates accurate segmentation from knee MR images. However, the meniscus is thin and has ambiguous boundaries, making segmentation difficult, and existing CNN- or Transformer-based methods often suffer from over- or under-segmentation. In this study, we propose Joint CondDiff, a conditional diffusion model that simultaneously noises both images and masks and employs step-wise image injection during sampling. Experimental results demonstrate improved precision and F1-score compared to SegDiff, with consistent performance gains even in challenging cases.