COMPARATIVE ANALYSIS OF 3D CNN, NNU-NET, AND SAM2-BASED MODELS FOR KNEE MRI SEGMENTATION

MRI enables assessment of morphological and compositional changes in knee cartilage and meniscus relevant to osteoarthritis (OA) progression. Accurate quantification depends on reliable segmentation, which remains challenging due to the thin, curved geometry of cartilage and complex shape of the meniscus. Manual delineation is time-consuming and impractical for large-scale studies. Deep learning models, including nnU-Net and transformer-based foundation models such as MedSAM2, have shown promise for automated segmentation; however, the effectiveness of fine-tuning MedSAM2 for multi-structure knee segmentation and its performance relative to nnU-Net on established benchmarks such as the IWOAI Challenge has not been reported.

1) To evaluate and compare the performance of domain-specific fine-tuned MedSAM2 and nnU-Net for automated multi-structure knee MRI segmentation using all OAI baseline cohort. 2) To compare the segmentation performance of both models on the IWOAI Knee Segmentation Challenge dataset.

Sagittal DESS knee MR images from the OAI baseline dataset (n = 9,592 MRI of left and right knees) and corresponding segmentation masks were used. Data was split at the subject level using stratified sampling across KL grades into 70/10/20 training, validation, and test sets. MedSAM2 was fine-tuned for 75 epochs, resampling sagittal slices to 512 × 512 resolution, using a combined Focal and Dice loss (20:1 weighting). Prompts were derived from the 25th, 50th, and 75th percentile slices of each 3D volume, with ground truth masks from these slices passed as initial prompts to propagate segmentation to the rest of the slices. For the IWOAI Challenge evaluation, the same data splits of the challenge (n = 120 training, 28 validation, 28 test) were used for fine-tuning both MedSAM2 and 3D nnU-Net. Performance was assessed using Dice score and IoU.

Using the entire OAI baseline dataset (Objective 1), KneeSAM2 achieved a mean Dice of 0.92 and IoU of 0.88, surpassing pretrained MedSAM2 by 0.20 in Dice and 0.23 in IoU. Improvements were most pronounced for smaller structures: patellar cartilage (Dice 0.90 vs. 0.41), meniscus (0.90 vs. 0.48), and tibial cartilage (0.89 vs. 0.64). Using the IWOAI Challenge dataset (Objective 2), MedSAM2-FT outperformed the published IWOAI benchmark for patellar cartilage (Dice 0.88 vs. 0.86) and matched it for femoral cartilage (0.90 vs. 0.90) but showed lower performance for tibial cartilage (0.75 vs. 0.89) and meniscus (0.85 vs. 0.88). The nnU-Net model, trained on the same IWOAI data splits, achieved the highest performance overall (Dice: femoral cartilage 0.95, tibial cartilage 0.95, patellar cartilage 0.93, meniscus 0.94), surpassing all other models.

Domain-specific fine-tuning improved MedSAM2 performance for segmenting small and anatomically complex knee structures, supporting the need for targeted adaptation of transformer-based foundation models (Objective 1). However, on the IWOAI Challenge benchmark (Objective 2), fine-tuned MedSAM2 showed competitive but variable performance across structures, whereas nnU-Net consistently achieved higher accuracy. These findings suggest that, under the current experimental settings and datasets, nnU-Net may offer stronger performance for multi-structure knee OA segmentation. Further improvements in prompting strategies, model adaptation, and training efficiency are needed to better leverage foundation models for specialized medical imaging tasks.