Shoulder Joint Image Segmentation Based on Joint Convolutional Neural Networks

Magnetic resonance imaging (MRI) is now commonly used for the examination and diagnosis of joints. A key step is to segment the bones of interest in MRI. This paper presents an algorithm for automatic segmentation of shoulder joint images based on a joint convolutional neural network model, which can accurately segment glenoid and humeral head in the shoulder image. This method includes two collaborative deep learning networks. The first network uses Mask R-CNN segmentation model to perform preliminary instance segmentation of glenoid and humeral head. The second network uses the probability maps of voxel belonging to the different objects (glenoid, humeral head, and background) as the constraint of the spatial location; thereby more accurate segmentation can be obtained. There are 50 groups of MRI which are used to train and test, the accuracy of Dice Coefficient, Positive Predicted Value (PPV), and Sensitivity for glenoid and humeral head reached 0.91±0.02, 0.95±0.01, 0.94±0.02 and 0.88±0.01, 0.91±0.02, 0.90±0.02 respectively, exceeding the current advanced segmentation algorithms.