3391 - Multi-Omics Models can Predict Prostate Specific Membrane Antigen Avidity for Computed Tomography Lesions in Oligo-Metastatic Castration Sensitive Prostate Cancer

R. Kumar¹, C. Zhang¹, P. Sutera², K. K. English², L. Hathout¹, S. K. Jabbour¹, L. Ren³, P. T. Tran³, R. Deek⁴, J. Kim¹, H. C. Onal⁵, O. C. Guler⁵, Y. Zhang¹, K. Nie¹, and M. Deek¹; ¹Department of Radiation Oncology, Rutgers Cancer Institute of New Jersey, New Brunswick, NJ, ²Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins Medicine, Baltimore, MD, ³Department of Radiation Oncology, University of Maryland School of Medicine, Baltimore, MD, ⁴Department of Biostatistics, University of Pittsburg, Pittsburg, PA, ⁵Department of Radiation Oncology, Baskent University Faculty of Medicine, Ankara, Turkey

Purpose/Objective(s): Prostate specific membrane antigen (PSMA) PET imaging is a sensitive modality for detection of early metastases in prostate cancer (PC). However, optimal integration of PSMA scans in routine follow-up is unknown. We hypothesized that radiomics combined with genomics and clinico-pathological information could predict PSMA avidity in lesions seen on conventional CT imaging allowing for more personalized patient selection. Materials/

Methods: Oligometastatic castrate-sensitive prostate cancer (omCSPC) patients with PSMA PET imaging, genomic mutational analysis and clinico-pathological details were included (n=84). Both PSMA avid and non-avid lesions in lymph node (LN) and bone were contoured manually on CT images. Radiomic features were extracted for all lesions using open source software. Various machine learning algorithms including support vector machine (SVM) and decision trees were used to predict PSMA avidity in CT anatomic correlates with five-fold cross validation. Descriptive statistics and receiver operating characteristics (ROC) curves were generated to report area under curve (AUC), sensitivity and specificity for each model.
Results: 1492 lesions were contoured, of which 234 lesions were PSMA avid, while 1258 lesions were non-avid. The PSMA avid/ total number of lesions in bone and LN were 47/332 (14.2%) and 164/1016 (16.14%), respectively. The median number of PSMA avid lesions per patient was 2 (IQR 1-3). Low volume disease (as per CHAARTED study) was seen in 96.4% patients. High-risk mutations (HiRi, defined as pathogenic mutations in ATM, BRCA1/2, Rb1 or TP53) were seen in 34.5% patients. Other pathogenic somatic mutations in the cohort were TP53 (20.2%), PTEN (14.3%), BRCA2 (8.3%), ATM (4.8%), BRCA1 (2.4%), CHEK2 (2.4%), ATRX (2.4%), PALB2 (2.4%), Rb1 (1.3%) and RAD51 (1.2%). After preprocessing, 107 radiomic features analyzing texture, shape and intensity were extracted per lesion. The top five important features were site of lesion, volume of lesion, total number of lesions seen on CT, Gray level non-uniformity and HiRi mutation status. Other important features selected were large area Gray level emphasis, coarseness, non-uniformity, CHEK2 mutation, FANCD2 mutation and pre-metastatic PSA. The best predictive model including radiomics, genomics and clinical factors was generated for detection of metastases on CT imaging. The overall AUC of the best model was 0.88-0.93 and can accurately predict 196/234 avid lesions (sensitivity of 83.7%), and 1138/1258 non-avid lesions (specificity of 90.4%).
Conclusion: Multi-omics including radiomics, genomics and clinical factors can predict PSMA avidity in early metastatic lesions in LN and bones on CT imaging in omCSPC. This approach can be applied for better personalization and decision-making regarding PSMA surveillance imaging. This multi-omics model needs to be further validated in a larger independent cohort.