Mp30-18 Machine-Learning Based Radiomic Models Predict Response to Neoadjuvant Chemotherapy in Upper Tract Urothelial Carcinoma

You have accessJournal of UrologyImaging/Uroradiology II (MP30)1 May 2024MP30-18 MACHINE-LEARNING BASED RADIOMIC MODELS PREDICT RESPONSE TO NEOADJUVANT CHEMOTHERAPY IN UPPER TRACT UROTHELIAL CARCINOMA Lennert Eismann, Mark Zucker, Burcin Agridag, Sam Keene, Armaan Kohli, Stephen Reese, Andreas Aulitzky, Mark Dawidek, Christian Stief, Abraham Ari Hakimi, Ed Reznik, and Jonathan Coleman Lennert EismannLennert Eismann , Mark ZuckerMark Zucker , Burcin AgridagBurcin Agridag , Sam KeeneSam Keene , Armaan KohliArmaan Kohli , Stephen ReeseStephen Reese , Andreas AulitzkyAndreas Aulitzky , Mark DawidekMark Dawidek , Christian StiefChristian Stief , Abraham Ari HakimiAbraham Ari Hakimi , Ed ReznikEd Reznik , and Jonathan ColemanJonathan Coleman View All Author Informationhttps://doi.org/10.1097/01.JU.0001009416.90901.7b.18AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: Neoadjuvant chemotherapy (NAC) prior nephroureterectomy is standard of care for patients with high-risk upper tract urothelial carcinoma (UTUC). However, response to NAC varies, and reliable predictive biomarkers remain elusive. Therefore, we evaluate different machine learning models to use radiomic features to predict response to NAC in patients with UTUC. METHODS: Our Institutional database was queried to identify patients receiving chemotherapy prior nephroureterectomy for UTUC between 2006 and 2022. Tumor segmentations were performed in contrast-enhanced CT scan by two independent annotators. Radiomic features (RF) were extracted, and intra-class correlation was assessed to evaluate concordance between annotators. Three machine-learning models were used to build a radiomic-model (RF only) to predict response to NAC. Secondly, we incorporated 15 pre-treatment clinical parameters to improve predictive accuracy (clinico-radiomic model). To estimate predictive value receiver operator curve (ROC) was used and summarized as area under the curve (AUC). RESULTS: In total, 148 patients were identified receiving NAC prior nephroureterectomy. 92 patients met inclusion criteria and tumor segmentation was performed. Pathological response to NAC was found in 58% (53/92). Using Pyradiomics, 1.690 RF were extracted, and high inter-reader concordance was found in 67% (1.132/1.690). 7 relevant clinical features (Figure 1) were identified and the 1.132 radiomics features were compressed to 10 dimensions using principal component analysis. Logistic regression (Logistic), support vector classifier (SVC) and gradient boosting (GB) model achieved an AUC of 0.64; 0.62 and 0.68 for RF only, respectively. The clinico-radiomic model reached an AUC of 0.65; 0.61 and 0.74. CONCLUSIONS: In this study, the majority of radiomic features were robust between annotators. Moreover, radiomic features correlated with response to NAC in UTUC. Machine-learning models incorporating radiomic features and clinical parameters show promise in predicting response, highlighting their potential clinical significance with further development. Download PPT Source of Funding: LE was supported by a postdoctoral fellowship of Deutsche Forschungsgemeinschaft (German Research Foundation) © 2024 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 211Issue 5SMay 2024Page: e501 Advertisement Copyright & Permissions© 2024 by American Urological Association Education and Research, Inc.Metrics Author Information Lennert Eismann More articles by this author Mark Zucker More articles by this author Burcin Agridag More articles by this author Sam Keene More articles by this author Armaan Kohli More articles by this author Stephen Reese More articles by this author Andreas Aulitzky More articles by this author Mark Dawidek More articles by this author Christian Stief More articles by this author Abraham Ari Hakimi More articles by this author Ed Reznik More articles by this author Jonathan Coleman More articles by this author Expand All Advertisement PDF downloadLoading ...