3219 - Apparent Diffusion Coefficient Tumor Volume is the Most Prognostic Radiomic Feature in Localized Prostate Cancer

J. E. Leeman¹, L. K. Lee², G. Pratt³, M. Rowan³, J. Parisi⁴, C. Belant⁴, K. N. Lee⁵, D. D. Yang⁵, P. F. Orio III⁶, P. L. Nguyen¹, A. V. DAmico³, and M. T. King¹; ¹Department of Radiation Oncology, Brigham and Women’s Hospital/Dana-Farber Cancer Institute, Boston, MA, ²Department of Radiology, Brigham and Womens Hospital, Boston, MA, ³Brigham and Womens Hospital/Dana-Farber Cancer Institute, Boston, MA, ⁴Dana-Farber Cancer Institute/Brigham and Womens Hospital, Boston, MA, ⁵Harvard Radiation Oncology Program, Boston, MA, ⁶Department of Radiation Oncology, Dana-Farber Cancer Institute, Boston, MA

Purpose/Objective(s): We have previously shown that tumor volume, defined as the volume of a deep learning segmentation result of the intraprostatic tumor on magnetic resonance imaging (MRI), was independently prognostic for oncologic outcomes after radiation therapy (RT) and radical prostatectomy (RP). However, other MRI features, such as mean apparent diffusion coefficient (ADC) intensity, are also prognostic. The purpose of this study is to evaluate whether other radiomic features may provide more prognostic information than tumor volume. Materials/

Methods: This is a single institutional retrospective review of 946 patients (573 RP, 373 RT), who had a PIRADS 3-5 lesion on mpMRI (GE Signa: 2010-2013; GE DISCOVERY: 2013-2017) prior to RP or RT. Reference segmentations were provided by GU radiation oncologists. Radiomic features were extracted from the ADC image utilizing PyRadiomics, and normalized. Patients were divided into 4 cohorts, based on treatment and mpMRI scanner type (A (N = 190): Signa/RT, B (183): Discovery/RT, C (172): Signa/RP, D (401): Discovery/RP). Within each cohort, patients were subdivided 2:1 into training and test sets. For each training set, 106 radiomic features were combined into 19 clusters utilizing k-medoids method. The mean of all included features was assigned to each cluster variable. The cluster most highly correlated with tumor volume was identified. Each cluster was then included as an independent variable in a univariable Cox regression model for estimating the time to biochemical failure. Individual clusters were ranked based on partial Cox likelihood across 1000 bootstrap resampling intervals. Clusters that were ranked higher than the tumor volume cluster were identified. Associations of such clusters with time to biochemical failure for the test set were estimated in multivariable models adjusted for age and NCCN risk category (low/favorable intermediate, unfavorable intermediate, high risk).
Results: For each cohort, ADC tumor volume was highly correlated (r>0.99) with a single cluster. Median ranks for the ADC tumor volume cluster across the 4 training cohorts were 2 [IQR: 2, 7] (A), 1 [1, 6] (B), 1 [1,5] (C), and 3 [2, 3] (D). Respective ranks for the cluster associated with mean ADC intensity were 13 [10, 16], 8 [5, 12], 16 [9, 17], and 7 [6, 10]. For test sets, ADC tumor volume was associated with biochemical failure across all cohorts (A: adjusted hazard ratio (AHR) 2.03 [1.14, 3.61]; p = 0.02), (B: 1.74 [1.19, 2.53]; p = 0.004), (C: 2.53 [1.05, 6.13]; p = 0.04), and (D: 1.92 [1.18, 3.12]; p = 0.009) in multivariable models. Other more highly ranked radiomic clusters were not significantly associated with biochemical progression.
Conclusion: ADC tumor volume is an independent, highly prognostic radiomic feature across scanner types and patients treated with either RT or RP. Other radiomic clusters are not significantly associated with biochemical progression.