MP44-18 ACCURATE ESTIMATE OF SPLIT DIFFERENTIAL RENAL FUNCTION USING ULTRASOUND ALONE FOR INFANTS WITH HYDRONEPHROSIS

Abstract

You have accessJournal of UrologyPediatric Urology II (MP44)1 Sep 2021MP44-18 ACCURATE ESTIMATE OF SPLIT DIFFERENTIAL RENAL FUNCTION USING ULTRASOUND ALONE FOR INFANTS WITH HYDRONEPHROSIS Marta Skreta, Lauren Erdman, Mandy Rickard, Daniel T. Keefe, Michael Chua, Joana Dos Santos, Anna Goldenberg, and Armando J. Lorenzo Marta SkretaMarta Skreta More articles by this author , Lauren ErdmanLauren Erdman More articles by this author , Mandy RickardMandy Rickard More articles by this author , Daniel T. KeefeDaniel T. Keefe More articles by this author , Michael ChuaMichael Chua More articles by this author , Joana Dos SantosJoana Dos Santos More articles by this author , Anna GoldenbergAnna Goldenberg More articles by this author , and Armando J. LorenzoArmando J. Lorenzo More articles by this author View All Author Informationhttps://doi.org/10.1097/JU.0000000000002065.18AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: Estimating differential renal function (DRF) requires imaging with nuclear renograms. Obtaining DRF from ultrasound (US) images would be beneficial to patients. Herein we explore applying our existing deep learning architecture to the task of determining DRF from ultrasound images. METHODS: We used serial US from 135 infants with hydronephrosis and data from lasix/DMSA renal scans. Function was classified as “normal” (>40;<60)/“abnormal” (<40; >60). We trained a convolutional neural network (CNN) with the task of predicting DRF from sagittal or transverse renal US images. We used US images of both the left and right kidneys and predicted the function for the left kidney, as DRF is measured in relation to both kidneys. We tested a CNN with 7 convolutional layers and 2 linear layers for single renal views and a neural network of two identical CNN subnetworks for assessment of 2 renal views. RESULTS: We were able to predict normal/abnormal renal function in US images with an AUROC 0f 0.776. Including both sagittal and transverse views in the model improved our performance (Table 1). To improve the interpretability of our predictions, we generated heat maps to view areas of interest in US images that our classifier deemed most indicative for predicting function abnormalities. CONCLUSIONS: Prediction of normal or abnormal function based on US images alone appears to be feasible without feature-engineering or clinical/patient variables. This technology may allow for closer monitoring of infants and reduce exposure to invasive testing patients receive to assess function. Source of Funding: none © 2021 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 206Issue Supplement 3September 2021Page: e799-e800 Advertisement Copyright & Permissions© 2021 by American Urological Association Education and Research, Inc.MetricsAuthor Information Marta Skreta More articles by this author Lauren Erdman More articles by this author Mandy Rickard More articles by this author Daniel T. Keefe More articles by this author Michael Chua More articles by this author Joana Dos Santos More articles by this author Anna Goldenberg More articles by this author Armando J. Lorenzo More articles by this author Expand All Advertisement Loading ...