Multiparametric MRI (mpMRI) including T2-weighted (T2W), diffusion weighted (DW) and dynamic contrast enhanced (DCE) imaging is the clinical standard for detection of primary prostate cancer [1]. Computer-assisted tumor detection algorithms, which typically involve a voxel classification step (e.g. [2]), have the potential to support radiological reading. The purpose of this study is to investigate if standardized uptake values (SUV) from combined ¹⁸F-Fluciclovine PET/mpMRI can improve voxel classification for computer-aided detection of localized primary prostate cancer.High-risk patients (Gleason score ≥ 8 and/or PSA > 20 and/or clinical stage ≥ cT3) scheduled for prostatectomy with extended lymph node dissection underwent a PET/MRI exam (3 T Biograph mMR, Siemens, Erlangen, Germany) prior to surgery, as part of an ongoing study to investigate the potential of combined ¹⁸F-Fluciclovine PET/mpMRI for staging of prostate cancer. The imaging protocol is specified in Table 1. DW images were corrected for distortion [3] after which apparent diffusion coefficient (ADC) maps were calculated using a mono-exponential model (including b=50/400/800 s/mm2). Volume transfer constant (K_trans) and extravascular extracellular volume fraction (V_e) maps were calculated from motion-corrected DCE images using the simple Toft’s model [4] with a population-based arterial input function [5]. PET counts summed over minutes 5 to 10 post ¹⁸F-Fluciclovine injection were reconstructed to SUV maps using a manufacturer-provided OP-OSEM reconstruction algorithm (4 iterations, 21 subsets, 6 mm FWHM Gaussian filter). All images were registered to the T2W image and resampled to the coarsest voxels size (DWI) using elastix [6]. Volumes-of-interest (VOIs) of tumor, nodular benign prostatic hyperplasia (BPH), stromal BPH, inflammation and healthy tissue were delineated on the T2-weighted images using OsiriX [7], having the corresponding whole-mount hematoxylin-eosin stained slides marked by a pathologist as a reference (Figure 1). The VOIs were also resampled to the DWI voxel size and each voxel was assigned to the tumor, healthy, benign lesion (BPH + inflammation), or non-tumor (healthy + benign lesion) class. Feature vectors consisting of T2W, ADC, K_trans, V_e and SUV voxel values were created. After Box-Cox transformation, statistical differences in means between classes were tested using linear regression with generalized estimation equations (GEE) to account for the assumed correlation between voxels of the same patient. Logistic regression with GEE was used to create 31 pet/mr and mr-only models for classification of tumor vs. healthy, tumor vs. benign lesion, and tumor vs. non-tumor voxels including data from all 24 patients. The mr-only and pet/mr model with the highest area under the receiver operating characteristic (AUROC) curve and with all feature regression coefficients significantly different from zero (p<0.05) were selected for leave-one-patient-out cross-validation. New ROC curves were obtained by training and testing the two selected models 24 times. Statistical differences between the AUROC curves were tested using Delong’s method. Matlab (Mathworks, Natick, MA, USA) was used for data analysis.Each feature vector consisted of 19250 voxels (1078 healthy, 6788 benign lesion, and 11384 tumor voxels from 37, 74, and 67 VOIs, respectively). All mean feature values were significantly different between tumor voxels and voxels from the other 3 classes (p<0.001) except for SUV between tumor and benign lesion voxels (p=0.1688). All MRI-derived features were selected for all three mr-only models, except for V_e in tumor vs. healthy voxels (Figure 2). SUV was additionally selected for the pet/mr models for classification of tumor vs. healthy and tumor vs. non-tumor voxels (Figure 2). However, the AUROC curves after cross-validation were not significantly different between the pet/mr and mr-only models and sensitivities at 90% specificity were similar (Figure3, Table 2).

¹⁸F-Fluciclovine PET is probably of most value for detection of prostate cancer recurrence following definitive therapy [8], but its merit for detection of localized primary disease in conjoint use with T2W MRI has also been described [9]. SUV was selected for tumor vs. non-tumor voxel classification, which is an important step in computer-assisted detection algorithms, but this did not lead to significant improvements in comparison with selecting only mpMRI-derived parameters. Overlap in SUV between tumor and benign lesion voxels is probably the most important reason. Whole-prostate analysis with sophisticated classifiers such as support vector machines might give better results and is subject of further research.

This initial analysis of combined ¹⁸F-Fluciclovine PET/mpMRI data suggests that SUV has limited additional value to mpMRI-derived features in voxel classification for computer-assisted detection of localized primary prostate cancer.