Benchmarking CNNs and Transformers for Knee Cartilage MRI Segmentation Without Pre-Training

 

Data

OAI subset of 176 DESS MRI scans from 88 patients (baseline and after one year) [1,2]

Scanner: Siemens 3T Trio Size: 384 x 384 x 160 volumes

Resolution: 0.36 x 0.36mm x 0.70mm

 

Models

CNN: nnU-Net [3]

Transformer: Segformer3D, SwinUNETR-V2 [4,5]

 

Evaluation Metrics

Volume: Dice similarity coefficient (DSC), volumetric overlap error (VOE)

Surface-distance: Hausdorff distance (HD95), average symmetric surface distance (ASSD)

Clinical: Average cartilage thickness error

 

Model Training

Training schedule Empirical parameters Fixed parameters Models trained on 3 NVIDIA L40S 48GB GPUs. nnU-Net training followed the original protocol with an ensemble of 2D and 3D models, and transformer-based model training protocol is shown in Figure 1.

 

 

Discussion, Conclusion and Further Work

• nnU-Net outperformed transformer based models and all IWOAI entrants, achieving SOTA accuracy for the dataset, but the performance difference was small

• In resource-limited settings without pre training, a well-designed CNNs remains the most effective choice for knee cartilage segmentation from MRI

• All models performed worse when tissue was significantly degenerate

• Test models on full OAI cohort (4,796 participants), linking clinical information to segmentation mask morphology

 

 

Acronyms

MRI = Magnetic Resonance Imaging

SOTA = State of the Art

OAI = Osteoarthritis Initiative

DESS = Double Echo Steady State

 

Acknowledgements

This work was funded by the UKRI Centre for Doctoral Training in Artificial Intelligence for Medical Diagnosis and Care at the University of Leeds [grant number EP/S024336/1/] and undertaken on the Aire HPC system at the University of Leeds, UK. The OAI dataset is publicly available. Ethical approval and informed consent were obtained by OAI for all participants.

 

 

References

[1] Branch, NIAMS Science Communications and Outreach. ‘Osteoarthritis Initiative’. National Institute of Arthritis and Musculoskeletal and Skin Diseases, NIAMS [2] Desai, Arjun D., et al. ‘The International Workshop on Osteoarthritis Imaging Knee MRI Segmentation Challenge: A Multi-Institute Evaluation and Analysis Framework on a Standardized Dataset’, Radiology: Artificial Intelligence

[3] Isensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: a self configuring method for deep learning-based biomedical image segmentation. Nat Methods. 2021 Feb;18(2):203–11.

[4] Perera S, Navard P, Yilmaz A. SegFormer3D: an Efficient Transformer for 3D Medical Image Segmentation [Internet]. arXiv; 2024 [cited 2025 Apr 17]. Available from: http://arxiv.org/abs/2404.10156

[5] He Y, Nath V, Yang D, Tang Y, Myronenko A, Xu D. SwinUNETR-V2: Stronger Swin Transformers with Stagewise Convolutions for 3D Medical Image Segmentation. In: Greenspan H, Madabhushi A, Mousavi P, Salcudean S, Duncan J, Syeda-Mahmood T, et al., editors. Medical Image Computing and Computer Assisted Intervention – MICCAI 2023. Cham: Springer Nature Switzerland; 2023. p. 416–26