0991
Clinical-genomic Risk Group Classification of Suspicions Lesions on Prostate mpMRI1University of Miami, Miami, FL, United States
Synopsis
The applications of prostate mpMRI in clinical decisions, related to the need for prostate biopsy and which areas to biopsy, have rapidly increased over the past few years. After the biopsy, clinicians also have series of methods to determine the aggressiveness of the prostate cancer. The National Comprehensive Cancer Network [NCCN]1 risk groups is one of the most commonly used system. The primary intent of the NCCN is to predict biochemical recurrance rather than survival outcomes such as distant metastasis (DM). Recently, NCCN was integrated with a genomic classifier, Decipher,2 optimized to predict the risk of DM. The resultant new 3-tier risk clinical-genomic classification (CGC) system grouped the patients in low-, intermidiate- and high-risk.3 Here we present radiomics-based approach to predict the low risk group based on the novel CGC.
Purpose
The applications of prostate mpMRI in clinical decisions, related to the need for prostate biopsy and which areas to biopsy, have rapidly increased over the past few years. After the biopsy, clinicians also have series of methods to determine the aggressiveness of prostate cancer. The National Comprehensive Cancer Network (NCCN)1 risk groups is one of the most commonly used system. The primary intent of the NCCN is to predict biochemical recurrence rather than survival outcomes such as distant metastasis (DM). Recently, NCCN was integrated with a genomic classifier, Decipher,2 optimized to predict the risk of DM. The resultant new 3-tier risk clinical-genomic classification (CGC) system groups the patients in low-, intermediate- and high-risk.3 Here we present radiomics-based approach to predict the low- and high-risk groups based on CGC.Methods
Data consisted of mpMRI datasets from 78 patients. The patients were enrolled in two institutional trails: an active surveillance trial: “MRI-Guided Biopsy Selection of Prostate Cancer Patients for Active Surveillance versus Treatment - MAST” (n=46) and a radiation treatment trial, “MRI-Guided Prostate Boosts Via Initial Lattice Stereotactic vs Daily Moderately Hypofractionated Radiotherapy – BlaStM” (n=32).
mpMRI data was acquired on 3T Discovery MR750 magnet (GE, Waukesha, WI) (n=42), 3T MR Magnetom Trio (n=5) and Skyra (29) and 1.5T Symphony (n=2) (Siemens, Erlagen, Germany). Biopsies were acquired on Uronav (InVivo, Gainsville, FL) MRI-ultrasound fusion instrument. Habitat Risk Score (HRS) maps4 were generated in MIM imaging platform (MIM Software, Cleveland, OH) for each patient. HRS is an quantitative imaging-based pixel by pixel 10 point score, optimized on prostatectomy Gleason Score (GS) and tumor volumes. To map the Region of interest (ROI), the volume of HRS6 in the area of the biopsy was used (Figure 1). Regions of normal appearing PZ (NAPZ) and TZ (NATZ) were contoured. The intensities of ROI, NAPZ and NATZ on T2-weighted (T2-w), Apparent Diffusion Coefficient (ADC), and BVAL – the high B-value image from the Diffusion Weighted Imaging (DWI-MRI), were characterized using nine histogram descriptors: 10%, 25%, 50%, 75%, 90%, mean, standard deviation, kurtosis, and skewness. Five texture features (energy, entropy, correlation, homogeneity, and contrast) were extracted also from T2-w, ADC and BVAL. The features were calculated using the grey level co-occurrence matrices (GLCM).5 The nine histogram descriptors, described above, were calculated for each texture feature.
Results
A total of 231 biopsies (median=3, range 1 – 10) with available Decipher prostate cancer classifier test (GenomeDx Biosciences, Vancouver, CA) were analyzed. CGC was calculated using NCCN and Decipher.3 The mpMRI exam was carried out on average 34 days before the biopsy.
We used the adaptive LASSO approach to predict the probability of each biopsy
being low- or high-risk. For each risk category we developed three models using: (i) patient's clinical characteristics (age, race, ethnicity, PSA, stage), (ii) radiomics variables and (iii) combined. The results are summarized in Table 1 and Table 2. Important
variables from the adaptive LASSO approach were selected based on importance
scores calculated by the bootstrap ranking procedure along with 10-fold
cross-validation and 200 bootstrapping repetitions.
Discussion
The results suggest that quantitative imaging features derived from mpMRI guided biopsies are associated with established clinical-pathologic characteristics. Further, radiomic features were correlated with known prognostic gene expression patterns in prostate cancer. extract features which have prognostic value and are more relevant for prostate cancer risk stratification. This radiogenomics approach has the potential to reduce overdiagnosis by more accurately detecting aggressive from indolent cancers over standard assessments of tumor burden (percent of tumor in each core and numbers of cores positive) and grade (Gleason score). The identification of imaging features that are associated with high grade and/or high volume disease and other habitats that are hardly ever or never yielding of such tumor characteristics on directed biopsies is a necessary first step to limiting biopsies only to regions of significance, thereby reducing (i) the potential morbidity of the procedure and (ii) the frequency of the procedure.Conclusions
Machine-learning radiomics-based model is capable of predicting novel clinical-genomic classification. While there are several reports of association of radiomics variables with GS or Grade Group (GG), to the best of our knowledge, this is the first report for predicting integrated clinical-genomic classification.Acknowledgements
This work was supported by National Cancer Institute [R01CA189295 and R01CA190105]References
1. Mohler JL, Armstrong AJ, Bahnson RR, et al. Prostate Cancer, Version 1.2016. J Natl Compr Canc Netw. 2016;14(1):19-30.
2. Erho N, Crisan A, Vergara IA, et al. Discovery and validation of a prostate cancer genomic classifier that predicts early metastasis following radical prostatectomy. PLoS One. 2013;8(6):e66855.
3. Spratt DE, Zhang J, Santiago-Jimenez M, et al. Development and Validation of a Novel Integrated Clinical-Genomic Risk Group Classification for Localized Prostate Cancer. J Clin Oncol. 2018;36(6):581-590.
4. Stoyanova R, Chinea F, Kwon D, et al. An Automated Multiparametric MRI Quantitative Imaging Prostate Habitat Risk Scoring System for Defining External Beam Radiation Therapy Boost Volumes. Int J Radiat Oncol Biol Phys. 2018.
5. Fehr D, Veeraraghavan H, Wibmer A, et al. Automatic classification of prostate cancer Gleason scores from multiparametric magnetic resonance images. Proc Natl Acad Sci U S A. 2015;112(46):E6265-6273.
6. Breiman L. Random forests. Mach Learn. 2001;45(1):5-32.
Figures
Figure 1. (left panel) Habitat Risk Score (HRS) maps in two prostate cancer patients. Prostate, peripheral zone (PZ) and urethra were manually contoured. (right panel) Corresponding volume of HRS6, used as biopsy ROI. Note that there are two biopsy ROIs in Patient 2.
