Deep learning radiomics based on Ultrasound images for assisted diagnosis of Chronic Kidney Disease

Abstract

Abstract Objective. This study aims to explore and discuss the application value of ultrasound images in chronic kidney disease(CKD) screening by constructing a CKD screening model based on gray-scale ultrasoundgraphs. Methods. According to the diagnostic criteria of chronic kidney disease, retrospective registration was conducted with patients who came to Tongde Hospital of Zhejiang Province. Specifically, renal ultrasound images of 110 patients with chronic kidney disease in various stages and 30 patients with non-chronic kidney disease were studied. A total of 1456 renal ultrasound images were thereby collected, including 296 normal kidney ultrasound images of non-chronic kidney disease, 193 kidney ultrasound images of CKD stage 1, 232 kidney ultrasound images of CKD stage 2, 429 kidney ultrasound images of CKD stage 3, 165 kidney ultrasound images of CKD stage 4, and 141 kidney ultrasound images of CKD stage 5. The data of each group are then divided into three sets in an 8:1:1 manner, namely training set with 1166 images (3496 images after data expansion), validation set with 146 images and testing set with 149 images. The convolutional neural networks are used for feature extractions and the screening model is constructed by fusing ResNet34 and texture features for recognizing CKD and its stage. The performance of the model is evaluated by the receiver operating characteristic curve (ROC). A comparison analysis is also carried out with comparing the diagnosis results from ultrasound medicine physicians of two different levels, namely expert-level associate chief physician and resident physician. Results. When diagnosing CKD or non-CKD based on renal ultrasound image,the accuracy, sensitivity, specificity and AUC of our model are 21.8%, 6.3%, 25.1% and 0.05 higher than those of expert physician group, and the diagnostic efficiency is higher than that of expert physician group, and the difference of AUC between our model and expert physician group is statistically significant,P < 0.05. In the stage diagnosis of CKD, the diagnostic sensitivity of our model in CKD G1-G3 is significantly higher than that of expert physician group, which increased by 51.7%, 56.8% and 21.9% respectively, and the diagnostic efficiency is also significantly higher than that of expert physician group. The difference of AUC is statistically significant,P < 0.05. The diagnostic efficiency of our model in CKD G4 and G5 can reach the level of expert physician group and there is no significant statistical difference between our model and expert physician group of AUC,P > 0.05. Conclusions. Our deep learning radiomics model based on grayscale ultrasound can obtain richer diagnostic information, which is more effective than expert level ultrasound physicians in the diagnosis of early chronic kidney disease and can assist in the early screening of chronic kidney disease.