Nora Alkhaldi
Chronic Kidney Disease (CKD) and Acute Kidney Injury (AKI) are serious health conditions, where ultrasound imaging serves as a non-invasive diagnostic tool. Diagnosis relies on identifying significant morphological features, such as a shrunken appearance in CKD or an enlarged, echogenic kidney in AKI. This study proposes and evaluates a two-stage deep learning framework developed using a clinical dataset collected from the Saudi Ministry of National Guard Health Affairs (NGHA). The process utilizes a modified U-Net model for precise kidney segmentation, followed by a novel dual-input classification stage that integrates both the original and segmented images to enhance feature extraction. A comprehensive analysis was conducted on four advanced architectures, including a Swin Transformer, DenseNet121, EfficientNetB0, and ResNet50. The results demonstrated that the ResNet50 model with dual-input configuration, preceded by SS-MUNet segmentation, was highly effective, achieving a test accuracy of 82.88% and an F1-score of 78.36%. Dual-input setups consistently outperformed single-input variants by 10-15% in F1-score across architectures, with ResNet50 and DenseNet121 showing the strongest generalization. This approach enhances diagnostic performance by using both segmented and non-segmented images to distinguish pathological kidney features with high accuracy. The findings establish a robust framework that holds considerable potential as a speedy, reliable, and accessible decision-support tool for clinical practice in nephrology. The proposed dual-input framework based on both original and segmented kidney images improves macro F1-score by an average of 12.5% in the range of 10.2–15.3% and accuracy by 9–13% compared to single-input baselines across four modern architectures on the independent test set. This clinically significant gain enables more reliable automated differentiation of AKI, CKD, and normal kidneys using only standard ultrasound.