Mp30-17 Radiomics-Based Multi-Class Classification for Composition of Urolithiasis on Non-Contrast Computed Tomography

You have accessJournal of UrologyImaging/Uroradiology II (MP30)1 May 2024MP30-17 RADIOMICS-BASED MULTI-CLASS CLASSIFICATION FOR COMPOSITION OF UROLITHIASIS ON NON-CONTRAST COMPUTED TOMOGRAPHY Seungwoo Jeong, Kwangsuk Lee, Jihun Lee, Lina Kim, and Hwiyoung Kim Seungwoo JeongSeungwoo Jeong , Kwangsuk LeeKwangsuk Lee , Jihun LeeJihun Lee , Lina KimLina Kim , and Hwiyoung KimHwiyoung Kim View All Author Informationhttps://doi.org/10.1097/01.JU.0001009416.90901.7b.17AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: To discover the composition of urolithiasis without any operations, we developed a machine learning model to predict Calcium Oxalate(CaOx), Calcium Oxalate+Calcium Phosphate(CaOx+CaPO), Uric acid, and Struvite of urolithiasis using non-contrast computed tomography (NCCT). METHODS: In total, 224 patients whose stones were removed surgically from 2009 to 2021 were used for training and validation. All patients underwent NCCT to confirm the location of stones before surgery, and discharged stones were analyzed to extract their composition by Fournier-transform infrared spectrometry. Using a pre-trained urinary stone segmentation model (Dice score 0.86 and IOU 0.74), stone masks were labeled in every slice of NCCT. Through postprocessing, stack the images in 3D, and divide the stones individually if there are multiple stones per patient. The masks were cross-checked by an experienced urologist, and manually corrected. Based on NCCT, Hounsfield units (HU) and radiomics features were extracted from 3D stacked images and 2D images with interquartile 25%, 50%, and 75% slices, respectively, depending on the location of the slices in the stone. A model for multi-class classification of four compositions using machine learning with HU and radiomics features was developed and tested. Figure 1. Patient cohort with more than 3 slices of stone for 2D Radiomics. RESULTS: The derivation set included a total of 276 stones from 224 patients (mean age 62.5, 64.7% for males) with Calcium Oxalate (86, 31.2%), Calcium Oxalate+Calcium Phosphate (108, 39.1%), Uric acid (41, 14.9%), Struvite (41, 14.9%). HU for each composition of CaOx (2D interquartile 25%/50%/75%/3D mean±SD: 1,026.1±108.0/1,199.3±111.8/930.7±88.0/994.4±45.8), CaOx+CaPO (1,178.5±124.1/1,296.4±126.1/974.5±106.4/1,029.0±59.8), Uric acid (449.8±44.0/477.0±41.4/422.7±39.5/402.0±20.1), Struvite (792.0±116.1/842.9±110.7/694.2±97.7/904.1±32.7). Machine learning-based multi-class classification model with LightGBM was evaluated by AUC, 2D interquartile 25%: 0.6498, 2D interquartile 50%: 0.8507, 2D interquartile 75%: 0.6289. CONCLUSIONS: Radiomics feature-based multi-class classification showed good performance in simultaneous prediction for urolithiasis composition. Download PPT Source of Funding: None © 2024 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 211Issue 5SMay 2024Page: e500 Advertisement Copyright & Permissions© 2024 by American Urological Association Education and Research, Inc.Metrics Author Information Seungwoo Jeong More articles by this author Kwangsuk Lee More articles by this author Jihun Lee More articles by this author Lina Kim More articles by this author Hwiyoung Kim More articles by this author Expand All Advertisement PDF downloadLoading ...