Classifying Chronic Kidney Disease in Ultrasound Images Using Deep Learning

Abstract

Ultrasonography is inexpensive and noninvasive compared to CT or MRI, but its limited range of motion, frequent artifacts, and dependence on operator experience make it very difficult for radiologists to accurately determine the criticality of chronic kidney disease (CKD). Therefore, the purpose of this study was to use deep learning to classify CKD of categories in ultrasound images with high accuracy. We used the right and left kidney long-axis tomogram, right and left kidney short-axis tomogram of 547 cases (G1: 75 cases, G2: 263 cases, G3: 186 cases, G4+5: 23 cases) who underwent renal ultrasound examination. We investigated seven deep learning models: AlexNet, GoogLeNet, ResNet18, ResNet50, ResNet101, Xception, and DenseNet201 with and without pretraining. Among the seven deep learning models, Xception with pretraining had the highest Area Under the Receiver Operating Characteristic Curve (AUROC): G1: 0.64±0.11, G2: 0.62±0.05, G3: 0.65±0.03, G4+G5: 0.60±0.08. In addition, the accuracy for each category in the test data was 65.1%. Since the number of cases in G1, G4, and G5 is very small compared to the number of cases in G2 and G3, we believe it is necessary to increase the number of training data by inverting or rotating the data vertically or horizontally in the future.