1Assistant Professor, Department of Artificial Intelligence and Data Science, V.S.B. Engineering College, Karur. Email: devinandhini1982@gmail.com
2Assistant Professor, Department of Artificial Intelligence and Data Science, V.S.B. Engineering College, Karur. Email: maalini.cse@gmail.com
3Final Year Student, Department of Artificial Intelligence and Data Science, V.S.B. Engineering College, Karur. Email: haripalanisamy13@gmail.com
4Final Year Student, Department of Artificial Intelligence and Data Science, V.S.B. Engineering College, Karur. Email: gokulPrasathvijayakumar@gmail.com
5Final Year Student, Department of Artificial Intelligence and Data Science, V.S.B. Engineering College, Karur. Email: yuvasudhan45@gmail.com
6Final Year Student, Department of Artificial Intelligence and Data Science, V.S.B. Engineering College, Karur. Email: yutha.r2004@gmail.com
Background: Having a proper detection of nephrolithiasis (Kidney Stone) in a medical image is one of the most prominent clinical challenges today. Whether detected initially or after the stone has become symptomatic, failure to diagnose the kidney stone disease at the earliest possible stage can lead to serious complications including renal obstruction, renal injury and systemic infection. This paper reports the advancement of an end-to-end Kidney Stone Detection System that employs a FASTER R-CNN, with a ResNet50-FPN-v2 backbone, for kidney stone detection and incorporates a Gradient-weighted Class Activation Mapping (Grad-CAM) module, along with an interactive Streamlit application for real world clinical use.
Methodology: The kidney stone detection model generates bounding box predictions of kidney stone regions (with each bounding box being assigned a per detection confidence score) in each image passed into the system, while the Grad-CAM layer highlights the image areas that contributed to each bounding box making the rational behind the model's prediction transparent to the clinician. Before sending images through the Kidney Stone Detection System, raw images passed through an Albumentations preprocessing pipeline where resizing, padding and normalising occurred resulting in each image being a 512 × 512 image prior to inference. The deployed Kidney Stone Detection System provides three operational modes. The first mode is Single Image Predict with an optional Grad-CAM heatmap overlay. The second mode is Batch Predict using all images in a directory with CSV reports and ZIP archives of annotated images available. The final mode is providing a Model Metrics dashboard that displays mean Average Precision scores using COCO standard.
Results: Results of the evaluation reveal an mAP@0.50 metric of 0.7706 indicating excellent (strong) spatial localization accuracy when detecting kidney stones together with high accuracy via the deep learning detector along with the incorporation of visual interpretability and a user-friendly clinical system.
Conclusion: The combination of these 3 systems provides an effective and scalable method for providing automated kidney stone screening in the hospital setting.
Keywords: Kidney Stone Detection, Faster R-CNN, Object Detection, Grad-CAM, Medical Image Analysis, Deep Learning, ResNet50 FPN v2, Streamlit Deployment, Explainable AI, Bounding Box Regression
How to cite this article: Nandhini Devi S, Maalini D, Hari P, Gokul Prasath V, Yuvasudhan S, Yutha R. Kidney Stone Detection Using Faster R-CNN with Grad-CAM Explainability and Interactive Web Deployment. Int J Drug Deliv Technol. 2026;16(13s): 200-204. DOI: 10.25258/ijddt.16.13s.21
Source of support: Nil.
Conflict of interest: None