Mp45-02 Evaluation of an Automated Ct-Based Deep Learning Image Segmentation Model for Stone Characterization in Difficult-to-Image Situations

You have accessJournal of UrologyStone Disease: Epidemiology & Evaluation II (MP45)1 May 2024MP45-02 EVALUATION OF AN AUTOMATED CT-BASED DEEP LEARNING IMAGE SEGMENTATION MODEL FOR STONE CHARACTERIZATION IN DIFFICULT-TO-IMAGE SITUATIONS Kimberly Maciolek, Joey Logan, Yong Fan, Gregory E. Tasian, and Ryan Hsi Kimberly MaciolekKimberly Maciolek , Joey LoganJoey Logan , Yong FanYong Fan , Gregory E. TasianGregory E. Tasian , and Ryan HsiRyan Hsi View All Author Informationhttps://doi.org/10.1097/01.JU.0001008764.86460.8e.02AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: CT-based stone volume is an important factor to evaluate stone burden but is difficult and time consuming to obtain. Automated methods for stone volume measurement are reproducible and accurate in most cases, however their performance may diminish in difficult-to-image situations. Here, we sought to investigate the performance of CT-based automated stone volume assessment in scenarios with concurrent indwelling tubes and surrounding hardware. METHODS: We retrospectively selected 52 CT scans obtained before and after percutaneous nephrolithotomy from 26 individuals with multiple co-morbidities including neurogenic bladder (n=13), spinal hardware (n=11), obesity (mean BMI 49 kg/m2, n=2), and limb contractures (n=2). Manually calculated stone volume was determined using an ellipsoid formula and summed for multiple stones. Automated stone volume measurements were performed using a previously validated deep learning image segmentation model for kidney parenchyma segmentation. Kidney stone volumes were calculated as outlier voxels with Hounsfield Unit (HU) greater than mean +4σ of HUs of all voxels within the kidney. RESULTS: Patients had a median age of 39 years (IQR 32-46 years) and were 46% male. Wide variability was noted in stone volume based upon manual and automated calculations (Figure 1). Multiple automated stone calculations could not be calculated (11/52, 21%) due to contrast (6/11, 55%) and spine hardware (5/11, 45%). CT scans with contrast resulted in less stone volume detected by the automated software than manually (median 60 vs 354 mm3, p=0.04). Although not statistically significant, less stone volume was detected with automated software for CT scans with indwelling stents or nephrostomy tubes (n=4, 8%, median 4,674 vs 7,446 mm3, p=0.28) and spinal hardware (n=12, 23%, median 0 vs 1007 mm3, p=0.15) had less stone volume detected by the automated software than manually. CONCLUSIONS: These findings motivate the development of CT-based automated stone detection algorithms that can account for IV contrast, indwelling tubes, and surrounding hardware, which are often seen in CT scans of high-risk stone formers. Download PPT Source of Funding: None © 2024 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 211Issue 5SMay 2024Page: e742 Advertisement Copyright & Permissions© 2024 by American Urological Association Education and Research, Inc.Metrics Author Information Kimberly Maciolek More articles by this author Joey Logan More articles by this author Yong Fan More articles by this author Gregory E. Tasian More articles by this author Ryan Hsi More articles by this author Expand All Advertisement PDF downloadLoading ...