Segmentation and classification of knee joint ultrasonic image via deep learning

The knee is one of the most complicated joints in the human body, but it could be easily injured. Ultrasound imaging is an important technology for the diagnosis of the knee disease. To assist doctors in the treatment and reduce errors of judgment, we investigate the segmentation of disease regions and the automated identification of the typical knee joint diseases. First, we use deep learning to segment the Region of Interest (ROI). To solve the mis-segmentation and poor edge segmentation that occur when the ultrasound image is directly fed into the deep neural network, an image segmentation framework is proposed that integrates snake preprocessing, dilated convolution to expand the receptive fields, and multi-channel learning. Second, due to the small difference in features among various categories of ultrasound images, a hybrid algorithm is proposed based on the Resnet rough classification and quadratic training with graph embedding. Finally, the experiments show that the proposed image segmentation framework achieves 10% greater accuracy than a common segmentation network. By visualizing the feature vectors extracted from the classification network, we verify that the feature vectors are closer on similar images after quadratic training by graph embedding. Employing the optimization with quadratic training, we increase the classification accuracy by 11% compared to the Resnet approach.