Pd36-11 a Novel Lightweight Machine Learning Model for Automated Reclassification of the Index Lesion on Biparametric Prostate Mri

You have accessJournal of UrologySurgical Technology ISUP Grade Group≥2). METHODS: Men who underwent 3T biparametric MRI (bpMRI) followed by MRI-informed targeted biopsy (TB) were identified (IRB# HS-13-00663). bpMRI scans were performed either in our institution or elsewhere according to Prostate Imaging-Reporting and Data System (PIRADS) v2 or v2.1 standards. The ROI (PIRADS 3-5 lesions) were manually segmented on thin (0.6mm) slice T2-weighted or standard T2-weighted, diffusion-weighted, and apparent diffusion coefficient, and were labeled according to CSPCa detection on TB. Index ROI was defined as the highest PIRADS and largest volume lesion. A lightweight and explainable machine learning model, the Green Learning (GL), was developed. The GL framework was meticulously designed by engineers, and physically defined imaging features only were selected for the feature extraction process; therefore, the GL was lightweight, explainable, and transparent. The classification performance for CSPCa on index ROI was analyzed by receiver operating characteristic (ROC) and logistic regression on the validation set. The optimal cutoff was determined by Youden's index. Statistically significant if p<0.05. RESULTS: A total of 474 patients with 253 thin bpMRI (tMRI) and 221 standard bpMRI (sMRI) were included. The GL was trained on randomly selected 317 bpMRI (169 tMRI and 148 sMRI) and validated on unseen 157 bpMRI (84 tMRI and 73 sMRI). The area under the ROC curve (AUC), accuracy, sensitivity, and specificity were: 0.79, 75%, 83%, and 71% for tMRI; 0.68, 66%, 66%, and 66% for sMRI, respectively. Using a GL probability cutoff of 0.28≥ for tMRI and 0.38≥ for sMRI, 82% and 77% of PIRADS 3 ROI without CSPCa could safely avoid unnecessary TB, achieving a negative predictive value of 96% and 100%, respectively. The GL probability score, PSA density≥0.15ng/mL2, and digital rectal examination abnormality were independent predictors of the CSPCa on TB. Integration of the predictors improved the GL AUC to 0.88 for tMRI and 0.80 for sMRI. CONCLUSIONS: We developed a novel Green Learning machine learning model to reclassify PIRADS ROI into CSPCa positive or negative, using multi-vendor MRI scanners. Combined with clinical predictors, the Green Learning offers a marked improvement over the current radiological PIRADS standard for CSPCa detection. Source of Funding: None © 2024 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 211Issue 5SMay 2024Page: e797 Advertisement Copyright & Permissions© 2024 by American Urological Association Education and Research, Inc.Metrics Author Information Giovanni E. Cacciamani More articles by this author Masatomo Kaneko More articles by this author Vasileios Magoulianitis More articles by this author Jiaxin Yang More articles by this author Yijing Yang More articles by this author Jintang Xue More articles by this author Jinyuan Liu More articles by this author Passant Mohamed More articles by this author Darryl H. Hwang More articles by this author Karanvir Gill More articles by this author Lorenzo Storino Ramacciotti More articles by this author Divyangi Paralkar More articles by this author Manju Aron More articles by this author Vinay Duddalwar More articles by this author Suzanne L. Palmer More articles by this author C.-C. Jay Kuo More articles by this author Chrysostomos L. Nikias More articles by this author Inderbir S. Gill More articles by this author Andre Luis Abreu More articles by this author Expand All Advertisement PDF downloadLoading ...