AUTOMATED QUANTIFICATION OF MENISCUS EXTRUSION IN MRI VIA AI FOUNDATION MODEL: PROOF OF CONCEPT USING A TRAINING-FREE FEW-SHOT SEGMENTATION APPROACH

Abstract

INTRODUCTION

Manual assessment of meniscus extrusion (ME) in magnetic resonance (MR) images is time-consuming and prone to variability, limiting efficiency in clinical and research settings. While deep learning methods have shown promise in MR image segmentation, their reliance on task-specific training and large annotated datasets limits scalability and adaptability.

OBJECTIVE

Building upon our previously developed AI foundation model, we aim to establish a fully automated pipeline for quantifying ME in knee MRI with our model training and eliminate the need for large annotated datasets.

METHODS

By providing a support set including a minimal number of segmentation examples, the AI Foundation Model enables accurate segmentation of knee anatomy and reliable ME measurement in a training-free, few-shot manner. In the study, we analyzed 3T MR images acquired using either T2-weighted or proton density MR sequences from 10 patients with mild osteoarthritis. Manual segmentations of femur, tibia, medial, and lateral menisci were performed by experts. Two patients, one with T2-weighted and one with proton density images, were randomly selected to build the support set. The remaining 8 patients comprised the testing set, which was used for both automated segmentation and model evaluation. Segmentation performance was assessed using the Dice Coefficient. For ME evaluation, an experienced radiologist manually identified the slice containing the tibial spine and measured extrusion as the reference. Automated ME measurement was computed from the segmentation by detecting the femoral condyle and tibial plateau edge, then measuring the distance from the most medial point of the medial meniscus to a reference line connecting the femoral condyle and tibial plateau edge.

RESULTS

The average Dice Coefficient was 94.07 ± 3.97% for the femur, 97.09 ± 0.93% for the tibia, 82.91 ± 6.72% for the medial meniscus, and 85.49 ± 5.24% for the lateral meniscus. ME measurements predicted by the model were also compared with ground truth values. The human measured ME was 4.26 ± 1.46 mm, while the model-predicted ME was 4.18 ± 1.16 mm.

CONCLUSION

This study demonstrates that the foundation model enables reliable and fully automated quantification of meniscus extrusion from knee MR images without requiring training or large annotated datasets. With only two support examples, the model achieved accurate segmentation and ME measurement across eight testing subjects, underscoring its efficiency and strong generalization. Its consistent performance across key anatomical structures highlights its potential for expert-level evaluation in both clinical and research settings with minimal manual effort. Further work will explore semi-automated expansion of the support set and extension to diverse MRI protocols and osteoarthritis severities, and validation on larger-scale datasets.