What question did this study set out to answer?

This study aimed to evaluate an AI model for detecting urinary stones using X-ray imaging in resource-constrained settings.

April 8, 2026Open Access

Artificial intelligence for radiographic detection of urinary stones in resource-constrained settings

Key Points

This study aimed to evaluate an AI model for detecting urinary stones using X-ray imaging in resource-constrained settings.
Conducted a retrospective observational study using radiographic images from two hospitals.
Collected a dataset of 11,900 radiographs, saving 2,410 images for analysis.
Annotated images using the Roboflow platform with collaboration from trained personnel.
Partitioned the dataset into training, validation, and test subsets for model training.
Developed and fine-tuned a YOLOv8-based object detection model for identification of urinary stones.
The model achieved an overall accuracy of 0.56, indicating modest performance.
Precision was recorded at 0.86, demonstrating a high rate of correctly identified stones among detected cases.
Recall was at 0.53, indicating a lower rate of detection among actual urinary stones.
Specificity was noted at 0.66, reflecting a reasonably good rate of correctly identifying non-stone cases.

Abstract

Urolithiasis is a common cause of emergency department visits, and early detection of urinary tract stones is critical to avoid complications, especially in healthcare systems with limited access to advanced imaging modalities. This study aimed to develop and internally evaluate an artificial intelligence (AI) model based on plain abdominal X-ray imaging for the detection of urinary stones in a resource-constrained setting. We conducted a retrospective observational study using radiographic images collected from two hospitals where X-ray is used as an initial diagnostic tool. From an initial collection of 11,900 radiographs, 2,410 images were saved. These images were annotated using the Roboflow platform by two trained study authors in collaboration with a radiology technician. The annotated dataset was partitioned into training (70%, n = 1,687), validation (20%, n = 482), and test (10%, n = 241) subsets, and a YOLOv8-based object detection model was fine-tuned for urinary stone identification. The model displayed modest overall performance, achieving an accuracy of 0.56 (259/459), precision of 0.86 (196/228), recall of 0.53 (196/364), and specificity of 0.66 (63/95). These results indicate that, while the current system does not yet yield high clinical profit in diagnostic settings, it demonstrates the technical feasibility of deploying AI for stone detection using widely available X-ray equipment. This work should be viewed as a proof of concept and a call to develop further and validate AI-based diagnostic tools that are designed to support existing imaging devices in resource-limited environments. Future research should focus on improving sensitivity, specificity, expanding and diversifying training datasets.

AI에게 질문

Bookmark

View Full Paper