Osteoarthritis (OA) is a degenerative joint disease and imposes an increasing burden on individuals and public health systems. Most prevalent joints are the knee, hip and hands, including the wrist. In order to enable early treatment of wrist OA, an early-detection method of cartilage loss, a characteristic symptom of OA, is needed. , CT images of the wrist bones. cannot visualize the cartilage itself, but instead use the distance between adjacent bones, to estimate the cartilage thickness. To enable such estimations, bones need to be segmented, which is a laborious task, that would impede any early diagnosis implementation. So automated segmentation of the wrist bones is desired for cost effective and objective assessments. However fully automatic segmentation of CT images is still a technical challenge. Deep learning techniques are considered a potentially successful approach to automate image processing.
The aim of this study is therefore to design and validate an automatic segmentation method of the capitate from CT-images based on a deep learning approach. For the automated segmentation method of CT images we selected UNet, a type of Convolutional Neural Network. A total of 10 CT images of the capitate, were divided into 3 groups to train (6), validate (2) and test (2) the network, while their corresponding segmented images were used as ground truth. Training and validation set were used during training to build the model, while test set was used after training to evaluate the performance of the model.
Quantitative evaluation of similarity between automatic segmentation by the network and the ground truth was expressed by the Dice coefficient (test data 1: 0.94, test data 2: 0.91) and the Hausdorff distance (test data1: 2.06mm, test data2: 2.55mm). Automatic segmentation took 6.7s for test data 1 and 8.1s for test data 2. The proposed approach holds promise for applications in fully automatic segmentation of wrist bones, as its performance, characterized by Dice coefficient and Hausdorff distance, is in par with those from other techniques of the same application. The next step in successful clinical implementation of the method is to improve the accuracy, for instance, by using a larger data size, after which the model can be further applied as an automatic quantitative metric in diagnosis of wrist OA.