Detection of kidney stone from ultrasound images using machine learning algorithms

Abstract

Nephrolithiasis is a prevalent cause of chronic renal diseases which is extremely costly to treat. The diagnosis of nephrolithiasis is difficult since there aren’t enough radiologist interpreters to interpret pictures from imaging devices and make a decision. Machine Learning (ML) algorithms are currently used for the detection or diagnosis of kidney stones, with the major drawbacks of limited data, ionizing radiation from scanning devices, ex-vivo techniques, and cost. In this research, ultrasound images are collected from different hospitals and annotated by radiographers or experts. Preprocessing mainly focused on filtering and segmentation for feature extraction and stone size estimation. Entropy and Gray Level Co-occurrence Matrix (GLCM) feature descriptors are extracted. In the analysis process, Support Vector Classifiers (SVC), Decision Trees (DT), K-Nearest Neighbors (KNN), and Random Forest (RF) algorithms are considered. KNN and RF models outperform the provided datasets. The KNN achieves performance metrics of accuracy, precision, recall, and AUC; 98.4%, 0.97, 1.0, and 0.98, respectively, and 95.1%, 0.94, 0.97, and 0.9896, respectively, for RF. Estimation of stone size with the major axis length of 10.2235 mm is obtained for the actual stone size of 11.9 mm, as annotated by the expert. Hence, the proposed approach of detecting kidney stones using ML algorithms can enhance and improve the diagnosis and detection of kidney stones (renal calculi) from ultrasound images, which are non-invasive, simple to use, and affordable without any ionizing radiation to improve the quality of life of the patients.