U-Net segmentation of the kidney on noncontrast CT for kidney stones location: a deep learning model for Mexican patients

This work presents the implementation of convolutional neural networks for the automated detection and segmentation of kidney stones in non-contrast computed tomography (CT) medical images from patients of the Mexican Institute of Social Security (IMSS). A U-Net architecture was employed to segment the kidneys and delineate regions of interest in high-resolution CT slices, with ground truth annotations provided by medical specialists. The model was trained and validated on a curated dataset representative of the Mexican population, achieving a Dice similarity coefficient of 0.9808 ± 0.038, an IoU of 0.9623 ± 0.0073, and a validation loss of 0.0040 ± 0.008 in five-fold crossvalidation. The proposed system demonstrated rapid convergence and excellent agreement with manual segmentations, confirming its utility as a reliable diagnostic aid for kidney stone detection in clinical practice.