Kidney Stone Detection

This study developed a deep learning system to automatically detect and measure kidney stone volume on low-dose, non-contrast CT scans, addressing limitations of prior non–deep learning approaches. Using a 3D U-Net for kidney segmentation followed by a CNN classifier, the model achieved strong detection performance (sensitivity 0.86 at 0.5 false positives per scan) and highly correlated volumetric measurements with manual standards (r² = 0.95). External validation on over 6,000 patients demonstrated excellent patient-level classification (AUC 0.95), confirming improved performance and generalizability compared with earlier methods.