Machine learning algorithms can recognize hydronephrosis in non-contrast CT images.

Abstract

BackgroundHydronephrosis, particularly attributed to the presence of renal calculi, is a clinical condition that can result in permanent renal injury, necessitating the utilization of imaging modalities for accurate diagnosis. Methodologies that can swiftly aid the radiologist by reducing workload are required for the preliminary diagnosis of hydronephrosis, which is critical in clinical practice.PurposeTo examine the efficacy of autosegmentation-assisted radiomics in predicting the presence of hydronephrosis among individuals diagnosed with renal colic.Material and MethodsThe study comprised 268 individuals who had non-contrast computed tomography (CT) scans presenting unilateral hydronephrosis. After the 3D autosegmentation of each patient's kidneys, first- and second-order radiomics parameters were acquired and Least Absolute Shrinkage and Selection Operator was employed as the dimensionality reduction tool. Machine learning (ML) procedures consisted of Support Vector Machine (SVM), Random Forest Classifier (RFC) analysis, Extreme Gradient Boosting (XGBoost), and Decision Tree Analysis.ResultsNo statistically significant difference was observed between the groups when comparing the side of hydronephrosis and the distribution of age among sexes. The repeated measurements of 3D autosegmentation exhibited a high level of intra-observer agreement. SVM, RFC, XGBoost, and Decision Tree analyses were able to predict the presence of hydronephrosis with AUC values of 0.966, 0.925, 0.994, and 0.978, respectively.ConclusionML-assisted radiomics can be considered an effective tool for accurately predicting the presence of hydronephrosis.