PD08-03 3D TEXTURE ANALYSIS IN ABDOMINAL CT AIDED BY MACHINE LEARNING CLASSIFIERS PREDICTS SHOCK WAVE LITHOTRIPSY SUCCESS

Abstract

You have accessJournal of UrologyStone Disease: Shock Wave Lithotripsy1 Apr 2018PD08-03 3D TEXTURE ANALYSIS IN ABDOMINAL CT AIDED BY MACHINE LEARNING CLASSIFIERS PREDICTS SHOCK WAVE LITHOTRIPSY SUCCESS Manoj Mannil, Jochen von Spiczak, Thomas Hermanns, Hatem Alkhadi, and Christian Fankhauser Manoj MannilManoj Mannil More articles by this author , Jochen von SpiczakJochen von Spiczak More articles by this author , Thomas HermannsThomas Hermanns More articles by this author , Hatem AlkhadiHatem Alkhadi More articles by this author , and Christian FankhauserChristian Fankhauser More articles by this author View All Author Informationhttps://doi.org/10.1016/j.juro.2018.02.449AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookTwitterLinked InEmail INTRODUCTION AND OBJECTIVES Predictive variables for successful shock wave lithotripsy (SWL) in patients with kidney stones include stone composition, size, shape, skin-to-stone distance (SSD) and mean calculus attenuation in Hounsfield Units (HU). However, CT attenuation values are influenced by differences in CT scan protocol settings, region-of-interest (ROI) delineation, magnification and windowing. Therefore published CT attenuation thresholds vary from 593 HU to 1350 HU and correlations between CT attenuation and success of SWL ranges considerably between studies. The aim of this study was to overcome the limitation of CT attenuation thresholds by applying 3D texture analysis (TA). METHODS CT was performed on patients with nephrolithiasis. All patients underwent SWL and received an abdominal CT examination before and after treatment and successful SWL was defined as residual stone size < 2 mm. 3D TA was performed on initial CT after postprocessing for pixel spacing and image normalization. We tested five commonly used machine learning-based models: (1) J48 decision tree (2) k-nearest neighbour (kNN), (3) artificial neural network (aNN) with backpropagation (Multilayer Perceptron), (4) Random Forest, and (5) sequential minimal optimization (SMO). In order to account for overfitting, the data set was split in the recommended ratio of 2/3 for model derivation and 1/3 for validation. Using C statistics we evaluated the discriminatory capability of uni- and multivariaable models with the 3D TA as compared with the models without 3D TA. RESULTS The model derivation and the validation cohort consisted of 34 and 17 patients with a median age of 55.5 and 53.9 years, 35% and 11% females, a median BMI of 27 and 28 and a median stone size of 10.8 and 9.1mm, respectively. Stone disintegration was successful in 22/51 (43.1 %) patients. In a multivariable model taking initial stone size, skin-to-stone distance, and BMI into account the Random Forrest represented the best predictive ability with a sensitivity of 0.59, specificity of 0.87 and an AUC of 0.81. CONCLUSIONS Our in-vivo study indicates the potential of TA to predict successful shock wave lithotripsy. © 2018FiguresReferencesRelatedDetails Volume 199Issue 4SApril 2018Page: e162 Advertisement Copyright & Permissions© 2018MetricsAuthor Information Manoj Mannil More articles by this author Jochen von Spiczak More articles by this author Thomas Hermanns More articles by this author Hatem Alkhadi More articles by this author Christian Fankhauser More articles by this author Expand All Advertisement Advertisement PDF downloadLoading ...