Abstract
<title>Abstract</title> <p>Chronic kidney disease (CKD) is a major global health burden, yet early diagnosis remains challenging and conventional kidney ultrasonography is limited by operator dependence and low sensitivity to subtle textural changes. We investigated the feasibility of an ultrasound foundation model for CKD classification from bilateral kidney ultrasound images, with reference labels defined according to the Kidney Disease: Improving Global Outcomes (KDIGO) 2024 guidelines. In this retrospective single-center study, 13,245 kidney ultrasound images from 1,484 patients (1,128 with CKD and 356 without) were analyzed. The Universal Ultrasound Foundation Model (USFM) was compared with five fine-tuned deep learning architectures (VGG16, ResNet-50, DenseNet-121, EfficientNet-B0, and Vision Transformer) over 10 random seeds, with performance assessed at the sample and examination levels. USFM outperformed the baseline models and showed greater robustness across repeated experiments. At the examination level, it achieved a balanced accuracy of 74.49 ± 4.63%, an F1-score of 73.59 ± 3.60%, and an area under the receiver operating characteristic curve of 84.80 ± 1.98%, with comparable results at the sample level. These findings demonstrate the potential of ultrasound foundation models to support CKD diagnosis from bilateral kidney ultrasound images.</p>