Neural network solutions based on U-Net architecture for automatic wrist joint cartilage segmentation in MR images

Segmentation of cartilage tissue in 3D magnetic resonance (MR) images is used to determine the stage of degenerative and inflammatory diseases of joints. For the wrist joint, manual segmentation is an extremely laborious task due to its complex structure. This determines the relevance of the development of fully automatic segmentation methods. The only automated method previously proposed is based on deep learning. It provided non-uniform segmentation accuracy depending on the slice position within the 3D image. This work aims to improve the accuracy of automatic segmentation of cartilage tissue in lateral slices of wrist joint MR images using deep convolutional neural networks (CNN). Two CNN architectures were considered: a classical U-Net architecture and a truncated version of U-Net, in which the deepest block of convolutions was removed. The segmentation accuracy was assessed using 3D and 2D Sørensen–Dice coefficients (DSC), as well as by calculating the area under the precision-recall curve (AUC-PR). The results were compared with previously published data for an automated method of cartilage segmentation of the wrist joint using a patch-based CNN, as well as with published results for a manual segmentation procedure. The use of U-Net-based architectures have significantly improved the automatic segmentation accuracy. The truncated U-Net architecture showed the best performance in terms of time (0.05 s per slice) and the highest segmentation accuracy (2D DSC = 0.77, AUC-PR = 0.844) among the reviewed CNNs for the test dataset of images. For slices without cartilage, the DSC increased from 0.21 to 0.75 using this architecture. Thus, the use of the U-Net architecture provided more uniform segmentation of 3D images than the method using the patch-based convolutional neural network.

1. Pap T., Korb-Pap A. Cartilage damage in osteoarthritis and rheumatoid arthritis—two unequal siblings. Nature Reviews Rheumatology. 2015;11(10):606–615. https://10.1038/nrrheum.2015.95

2. Moser T., Dosch J., Moussaoui A., Buy X., Gangi A., Dietemann J. Multidetector CT arthrography of the wrist joint: How to do it. RadioGraphics. 2008;28(3):787–800. DOI:10.1148/rg.283075087

3. Li A.E., Lee S.K., Rancy S.K., Burge A.J., Potter H.G., Wolfe S.W. Comparison of magnetic resonance imaging and radiographs for evaluation of carpal osteoarthritis. Journal of Wrist Surgery. 2017;6(2):120-125. DOI:10.1055/s-0036-1592140

4. Zink J.V., Souteyrand P., Guis S., Chagnaud C., Fur Y.L., Militianu D., et al. Standardized quantitative measurements of wrist cartilage in healthy humans using 3T magnetic resonance imaging. World journal of orthopedics. 2015;6(8):641-648. https://doi:10.5312/wjo.v6.i8.641

5. Peterfy C.G., van Dijke C.F., Lu Y., Nguyen A., Connick T.J., Kneeland J.B., et al. Quantification of the volume of articular cartilage in the metacarpophalangeal joints of the hand: accuracy and precision of three-dimensional MR imaging. American Journal of Roentgenology. 1995;165(2):371–375. https://10.2214/ajr.165.2.7618560

6. Dice L.R. Measures of the Amount of Ecologic Association Between Species. Ecology. 1945;26(3):297-302.

7. Prasoon A., Petersen K., Igel C., Lauze F., Dam E., Nielsen M. Deep feature learning for knee cartilage segmentation using a triplanar convolutional neural network. Advanced Information Systems Engineering. 2013;16(2):246-53. https://10.1007/978-3-642-40763-5_31

8. Panfilov E., Tiulpin A., Juntunen M., Casula V., Nieminen M., Saarakkala S. Automatic knee cartilage and menisci segmentation from 3D-DESS MRI using deep semi-supervised learning. Osteoarthritis and Cartilage. 2019;27(1):390–391. DOI:10.1016/j.joca.2019.02.391

9. Brui E., Efimtcev A.Y., Fokin V.A., Fernandez R., Levchuk A.G., Ogier A.C., et al. Deep learning‐based fully automatic segmentation of wrist cartilage in MR images. NMR in Biomedicine. 2020;33(8):1-13. DOI:10.1002/nbm.4320

10. Ronneberger O., Fischer P., Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. MICCAI 2015. Lecture Notes in Computer Science. 2015;9351:234-241. DOI:10.1007/978-3-319-24574-4_28