Knee cartilage segmentation on MRI using an expert-annotated clinical dataset

Knee osteoarthritis (OA) is a prevalent musculoskeletal disorder, where accurate segmentation of knee cartilage on magnetic resonance imaging (MRI) is pivotal for early diagnosis and treatment planning. Despite advances in deep learning, current methods often struggle with the precise segmentation of thin, irregular cartilage structures. Furthermore, the lack of high-quality, publicly annotated datasets, particularly those specific to Asian populations, hinders clinical translation. Therefore, the aim of this study is to construct a comprehensively annotated clinical dataset and develop an optimized Transformer-based framework to achieve precise, automated multi-class segmentation of the knee osteochondral unit. We constructed a comprehensively annotated Osteoarthritis MRI (OAMRI) dataset, comprising 893 T2-weighted sagittal knee MRI images from 47 Chinese patients, with the assistance of orthopedic experts from the Fourth Medical Center of Chinese PLA General Hospital.This dataset includes detailed annotations for four key anatomical structures: the femur, femoral cartilage, tibia, and tibial cartilage.Based on this dataset, we tailored and optimized the Swin-Unet baseline for knee structures by evaluating the cross-entropy loss function and employing the Adam optimizer.We validated our model on both the proprietary OAMRI dataset and the public MICCAI SKI10 benchmark. Extensive experiments demonstrated that our optimized model outperformed state-of-the-art methods on both datasets. Initially, under an exploratory slice-level split, the model achieved a theoretical upper-bound DSC of 95.45% and 95.84% for femoral and tibial cartilages. More importantly, when evaluated under a strict patient-level split to ensure true clinical generalization, the model maintained robust performance, achieving a DSC of 90.62% for femoral cartilage and 89.85% for tibial cartilage (averaging 91.43% for overall structures). On the SKI10 dataset, the model achieved DSC scores of 90.36% and 93.98%. The proposed optimized Swin-Unet model, coupled with the openly available OAMRI dataset, provides a powerful tool for precise knee cartilage segmentation. This approach holds significant potential for facilitating quantitative OA severity assessment and supports clinical decision-making for cartilage repair and replacement therapies.