Abstract: PD65-08
Sources of Funding: Research reported in this publication was supported by the National Cancer Institute of the National Institutes of Health under award numbers _x000D_ R21CA179327-01, _x000D_ R21CA195152-01, _x000D_ U24CA199374-01_x000D_ the National Institute of Diabetes and Digestive and Kidney Diseases under award number R01DK098503-02, _x000D_ the DOD Prostate Cancer Synergistic Idea Development Award (PC120857); _x000D_ the DOD Lung Cancer Idea Development New Investigator Award (LC130463),_x000D_ the DOD Prostate Cancer Idea Development Award; _x000D_ the Case Comprehensive Cancer Center Pilot Grant_x000D_ the VelaSano Grant from the Cleveland Clinic_x000D_ the Wallace H. Coulter Foundation Program in the Department of Biomedical Engineering at Case Western Reserve University_x000D_ the I-Corps@Ohio Program_x000D_ Case Urology Translational Research Training Program (CUTRTP)_x000D_ Hartwell Foundation_x000D_

Introduction

Multi-parametric magnetic resonance imaging (mp-MRI) based prostate imaging reporting and data system (PIRADS) is limited in confidently and robustly distinguishing clinically significant and insignificant prostate cancer (PCa). Radiomic features employ image processing methods to characterize specific patterns in images and have been shown to better characterize PCa than mp-MRI signal intensities alone. For example, gradient features quantify the appearance of edges, Haralick features distinguish homogenous low intensity (PCa) from normal regions and Gabor features quantify appearance of PCa at multiple orientations and scales. In this study, we aim to identify which of the mp-MRI derived radiomic features can distinguish high and low risk PCa as defined by the D’Amico criteria.

Methods

A retrospective cohort of 452 PCa patients who underwent a 3 Tesla mp-MRI scan was considered for this study. A subset of 72 patients comprising 153 lesions was chosen chronologically based on PIRADS to obtain a statistically balanced cohort. D’Amico criteria were available for 83 lesions and was used to categorize into low (N= 26), intermediate (N = 43) and high (N = 14) risk groups. A balanced dataset of N = 28 lesions with 14 lesions from each of high and low risk categories was finally assembled for radiomic feature analysis.

Results

A set of 101 radiomic features were extracted on a voxel-wise basis within the lesion region of interest (ROI) from each of T2w and ADC MRI sequences. First order statistics (mean, variance, skewness and kurtosis) were computed within each ROI to obtain 808 features per ROI. Of these, 44 features showed statistically significant differences between high and low risk lesions. Specifically, variance and skewness of T2w gradient and Gabor features, skewness and kurtosis of ADC Haralick and Laws features showed p<0.05 using Wilcoxon Rank-Sum test (representative results are shown in Figure). A random forests classifier trained using these radiomic features within a 3-fold cross validation framework resulted in an AUC of 0.96.

Conclusions

Radiomic features derived from mp-MRI distinguish high and low risk prostate cancer lesions as defined by D’Amico criteria. An independent validation of these features is required on a separate test set.

Funding

Research reported in this publication was supported by the National Cancer Institute of the National Institutes of Health under award numbers _x000D_ R21CA179327-01, _x000D_ R21CA195152-01, _x000D_ U24CA199374-01_x000D_ the National Institute of Diabetes and Digestive and Kidney Diseases under award number R01DK098503-02, _x000D_ the DOD Prostate Cancer Synergistic Idea Development Award (PC120857); _x000D_ the DOD Lung Cancer Idea Development New Investigator Award (LC130463),_x000D_ the DOD Prostate Cancer Idea Development Award; _x000D_ the Case Comprehensive Cancer Center Pilot Grant_x000D_ the VelaSano Grant from the Cleveland Clinic_x000D_ the Wallace H. Coulter Foundation Program in the Department of Biomedical Engineering at Case Western Reserve University_x000D_ the I-Corps@Ohio Program_x000D_ Case Urology Translational Research Training Program (CUTRTP)_x000D_ Hartwell Foundation_x000D_