A grading method for Kayser Fleischer ring images based on ResNet.

The corneal K-F ring is the most common ophthalmic manifestation of WD patients. Early diagnosis and treatment have an important impact on the patient's condition. K-F ring is one of the gold standards for the diagnosis of WD disease. Therefore, this paper mainly focused on the detection and grading of the K-F ring. The aim of this study is three-fold. Firstly, to create a meaningful database, the K-F ring images are collected which contains 1850 images with 399 different WD patients, and then this paper uses the chi-square test and Friedman test to analyze the statistical significance. Subsequently, the all collected images were graded and labeled with an appropriate treatment approach, as a result, these images could be used to detect the corneal through the YOLO. After the detection of corneal, image segmentation was realized in batches. Finally, in this paper, different deep convolutional neural networks (VGG, ResNet, and DenseNet) were used to realize the grading of the K-F ring images in the KFID. Experimental results reveal that the entire pre-trained models obtain excellent performance. The global accuracies achieved by the six models i.e., VGG-16, VGG-19, ResNet18, ResNet34, ResNet50, and DenseNet are 89.88%, 91.89%, 94.18%, 95.31%, 93.59%, and 94.58% respectively. ResNet34 displayed the highest recall, specificity, and F1-score of 95.23%, 96.99%, and 95.23%. DenseNet showed the best precision of 95.66%. As such, the findings are encouraging, demonstrating the effectiveness of ResNet in the automatic grading of the K-F ring. Moreover, it provides effective help for the clinical diagnosis of HLD.