LoRA-Based Multimodal Foundation Model Adaptation for MR Meniscus Segmentation

In this paper, we propose a method for stable meniscus segmentation in knee MR images under data-limited settings by combining a multimodal foundation model pre-trained on large-scale abdominal CT data with a parameter-efficient fine-tuning technique, LoRA.The proposed method jointly exploits image information and text prompts encoding anatomical semantics, and applies LoRA to selected layers of the image and text modules to improve both computational efficiency and segmentation performance.Experimental results demonstrate that the proposed model achieves high segmentation accuracy for both lateral and medial menisci while updating only a subset of the total parameters, and maintains stable performance even with a limited number of training samples.In addition, by systematically analyzing the segmentation performance according to the placement of LoRA modules, we identify an efficient LoRA adaptation strategy tailored for meniscus segmentation.