Abstract
AI education and research: examples of proof of concept or AI in development, technical advances, teaching approaches or preclinical testingKnee osteoarthritis (OA) is a multi-tissue joint condition that leads to chronic pain and joint stiffness, often characterised by cartilage loss.Magnetic resonance imaging (MRI) enables detailed assessment of knee OA, and segmentation of knee cartilage supports analyses of cartilage shape and thickness.Automatic segmentation is less laborious, less time-consuming and more reproducible compared with manual approaches.Convolutional neural networks (CNNs) have had significant success in automatic medical image segmentation; however, recent advances have focused on transformers.Transformers have demonstrated state-of-the-art performance in several domains but often require extensive pre-training, which is resourceintensive.Directly comparing CNNs and transformers trained with identical supervised conditions and resources provides insight into their relative strengths and supports model selection decisions, yet few studies have examined this for knee cartilage segmentation.This study evaluated CNN and transformer models, trained under identical supervised conditions, when segmenting four types of knee cartilage from MRI: the femoral, tibial and patellar cartilage and the meniscus.The MRI data set was from the Osteoarthritis Initiative and contained 176 scans from 88 participants.Each scan was collected from a Siemens 3T Trio MRI scanner using a double-echo steady-state sequence resulting in 384 x 384 x 160 volumes with image resolution of 0.36 mm x 0.36 mm x 0.70 mm.Scans were split into training, validation and testing sets of 120, 28 and 28 scans, respectively.Ground truth masks were generated by a single expert at Stryker Imorphics.Models included nnU-Net (CNN), SwinUNETR-V2 and Segformer3D (transformer).Performance was assessed using Dice score, volumetric overlap error, Hausdorff distance, average symmetric surface distance and average thickness error.For all tissues on most measures, nnU-Net outperformed transformer-based models, achieving Dice scores of 0.91±0.02,0.90±0.03,0.87±0.08 and 0.89 ±0.03 for the femoral, tibial and patellar cartilage and meniscus respectively.SwinUNETR-V2 achieved similar performance, with Dice scores of 0.90±0.02,0.88±0.03,0.85±0.07,0.87±0.03,while Segformer3D achieved significantly lower Dice scores of 0.81±0.04,0.81±0.06,0.73±0.09,0.80±0.03.These findings indicate that in resource-limited settings without pre-training, CNNs remain the most effective choice for MRI knee cartilage segmentation.While transformers can approach CNN performance, their reliance on pretraining may limit practicality.However, the small gap between top-performing models highlights opportunities for optimised architectures.