18F-Fluoroethylcholine PET/CT Radiomic Analysis for Newly Diagnosed Prostate Cancer Patients: A Monocentric Study

Aim: The aim of this study is to assess whether there are some correlations between radiomics and baseline clinical-biological data of prostate cancer (PC) patients using Fluorine-18 Fluoroethylcholine (18F-FECh) PET/CT. Methods: Digital rectal examination results (DRE), Prostate-Specific Antigen (PSA) serum levels, and bioptical-Gleason Score (GS) were retrospectively collected in newly diagnosed PC patients and considered as outcomes of PC. Thereafter, Volumes of interest (VOI) encompassing the prostate of each patient were drawn to extract conventional and radiomic PET features. Radiomic bivariate models were set up using the most statistically relevant features and then trained/tested with a cross-fold validation test. The best bivariate models were expressed by mean and standard deviation to the normal area under the receiver operating characteristic curves (mAUC, sdAUC). Results: Semiquantitative and radiomic analyses were performed on 67 consecutive patients. tSUVmean and tSkewness were significant DRE predictors at univariate analysis (OR 1.52 [1.01; 2.29], p = 0.047; OR 0.21 [0.07; 0.65], p = 0.007, respectively); moreover, tKurtosis was an independent DRE predictor at multivariate analysis (OR 0.64 [0.42; 0.96], p = 0.03) Among the most relevant bivariate models, szm_2.5D.z.entr + cm.clust.tend was a predictor of PSA levels (mAUC 0.83 ± 0.19); stat.kurt + stat.entropy predicted DRE (mAUC 0.79 ± 0.10); cm.info.corr.1 + szm_2.5D.szhge predicted GS (mAUC 0.78 ± 0.16). Conclusions: tSUVmean, tSkewness, and tKurtosis were predictors of DRE results only, while none of the PET parameters predicted PSA or GS significantly; 18F-FECh PET/CT radiomic models should be tested in larger cohort studies of newly diagnosed PC patients.


Introduction
Prostate cancer (PC) is the most frequent type of cancer in men worldwide. PC patients may be treated with radical prostatectomy or radiotherapy as primary treatments, but they may develop disease recurrence. Although the 10-year survival rate of PC patients is approximately 90% [1], 20-40% of patients will relapse [2][3][4]. Risk classification of patients with PC mainly relies on clinical examination, biochemical data, and Gleason Score (GS) [5], the latter being restricted to the biopsied portion of the prostate [6].
Positron Emission Tomography (PET) combined with Computed Tomography (PET/ CT) or Magnetic Resonance (PET/MR) using several radiopharmaceuticals are newgeneration imaging techniques allowing a more accurate and precise staging and restaging of PC compared to conventional imaging methods due to the advantage of combining functional and morphological information [7]. Radiolabelled choline is a radiopharmaceutical used in PET imaging as a marker of cell membrane synthesis, which is increased in PC. In particular, Fluorine-18 Fluoroethylcholine ( 18 F-FECh) PET/CT is a well-established non-invasive tool for restaging and staging/restaging settings of PC, with a reported patient-based overall sensitivity of 93% and a specificity of 83% [8,9].
Radiomics is a high-throughput approach that translates medical images into minable data by extracting many quantitative features describing the intensity, shape, and heterogeneity of targeted lesions. Radiomic models could indirectly be an expression of tumor biology behavior and could be considered an additional support, together with clinical and histological data, to help the clinicians in the diagnostic iter [10]. Radiomics applied to 18 F-FECh PET/CT was recently shown to predict PC outcome at follow-up [11]. However, the association between 18 F-FECh PET/CT radiomic features and initial tumor characteristics that may underlie these results has not been well described so far. Our study aimed at correlating the semiquantitative and radiomic parameters extracted from the prostate showing high 18 F-FECh uptake with baseline clinical assessment and histological data in patients undergoing PET/CT scans for initial staging.

Conventional Semiquantitative PET Parameters
This single-center retrospective study explored the association of semiquantitative PET parameters and radiomic features derived from primary tumor 18 F-FECh uptake with baseline clinical-biological data of patients with newly diagnosed PC.
None of the conventional PET parameters was a predictor of PSA values higher than PSA median value (9.3 ng/mL) or GS > 7. Volume, as explored by tMTV, was not associated with DRE positivity, while prostatic tKurtosis was, even at multivariate analysis. Our main hypothesis is that kurtosis, which represents a histogram feature and describes the shape of the distribution of 18 F-FECh values within the prostate affected by the tumor, conveys some sense of firmness or irregularity of the prostatic gland. Although not clinically relevant, this result may improve our understanding of the relationship between radiomic factors and the physical and biological properties of prostatic pathological tissues.
The role of semiquantitative 18 F-FECh PET/CT parameters is controversial in initial PC staging. Zanoni et al. showed a significant correlation between TBR and the International Society of Urological Pathology Score (derived from GS), using the liver as the reference organ [12]. These data are discrepant with our results, at least partially owing to the fact that we used the spleen as the reference organ. This choice was originally justified by the prevalence of metastases in the liver (about 10% [13], while they are very unusual in the spleen) and their detection difficulty on 18 F -FECh PET/CT [14].
Semiquantitative PET parameters were revealed to be useful for significantly discriminate malignant vs. non-malignant lesions [12]. Schaefferkotter et al. showed that SUV could be a strong predictor of aggressive disease in the prostate (GS ≥ 4 + 3) [15]. Conversely, semiquantitative analysis would not be able to discriminate among less-aggressive PC or benign conditions, such as foci of prostatitis and benign prostatic hyperplasia, because all of them are characterized by high tracer uptake due to the increased choline transport and overexpression of choline kinase in all these conditions [16].

Advanced Radiomic Features
The analysis of advanced radiomic parameters provided richer results, as a bivariate model with an accuracy of about 80% was found for each of the 3 outcomes: szm_2.5D.z. entr + cm.clust.tend for PSA levels (mAUC 0.83 ± 0.19), stat.kurt + stat.entropy for DRE (mAUC 0.79 ± 0.10), and cm.info. corr.1 + szm_2.5D.szhge for GS (mAUC 0.78 ± 0.16). The physical meaning of these features is less easily conceptualized, however, with several measures of heterogeneity (entropy). The association of textural parameters and GS opens interesting perspectives, as PET provides information on the whole prostate. The heterogeneous texture might not concern only the high uni-or multifocal 18 F-FECh uptake areas but could involve the entire gland (Figures 1 and 2). In line with this concept, Tu et al. assessed the predictive role of 3 radiomic zones within the prostate (zone 1 corresponding to the high metabolic volume with 40-100% of SUVmax value; zone 2, corresponding to the peripheral tumor zone with 30-40% SUVmax value; zone 3, corresponding to the extended peripheral tumor zone in which the entire prostate organ and zone 2 is included) for risk classification of PC patients referred to Choline PET/MR. The authors found that the zone 1 radiomics could be superior to PSA-based evaluation for patients' risk classification; moreover, both zone 2 and zone 3 were shown to be superior to TNM-based assessment for risk classification [17]. In our cohort, neither conventional PET parameters nor advanced radiomic models were able to discriminate significantly between low and high GS tumors. Our findings were different from the results drawn by Zamboglou et al., which showed the feasibility of using 68 Ga-PSMA radiomic features to discriminate GS 7 and GS 8 lesions [18]. Such discrepancy could at least partially rely on the different radiotracers and different scanners. Despite the role of radiomic models applied to the 18 F-FECh PET/CT images in PC patients that are emerging with the aim to provide support for clinical decisions in this setting in the era of personalized medicine, risk classification is still based on biopsy-based GS, which is not always representative of the whole tumor and may lead to side effects, such as hematuria and/or hematospermia, in addition to the trend of under-or over-grading primary PC [1]. Hence, imaging is still used for displaying suspected lesions and/or for biopsy guidance [19,20].      However, it should be kept in mind that 18 F-FECh PET/CT remains a well-established second-line diagnostic tool for clinical management of PC patients, with PSA values representing the main weapon for screening men of a certain age or with symptoms suspected of PC, such as dysuria. Furthermore, collecting a small amount of peripheral blood to assess PSA serum values is indisputably more convenient and feasible than the 18 F-FECh PET/CT procedure in terms of radiation exposure or medical expense. In summary, 18 F-FECh PET/CT represents a support, not an alternative tool for clinical baseline decision-making in PC patients, solidly based on GS and PSA values.

Limitations and Solutions
The main limitation of our study is its retrospective nature and, in particular, the heterogeneity of the PET/CT scanners in terms of hardware (digital vs. non-digital PET) and software (reconstruction, time of flight). Nevertheless, cross-fold validation ensured that the identified parameters were at least somewhat robust across the three scanners.
The number of included patients is overall low, but enough to derive significant models. Confirmatory studies should also be conducted to provide an external validation cohort for the confirmation of the observed results.
This work was confined to the study of the association between PET and clinicalbiological variables without any prognostic aim. Furthermore, our results are limited as they take into account the lack of definitive histopathological results of the whole prostate specimen.
Other studies may be needed to explore survival outcomes.

Materials and Methods
Patients with biopsy-confirmed PC who underwent a staging 18 F-FECh PET/CT in the PET/CT Center of Fondazione Policlinico Universitario A. Gemelli IRCCS (Rome, Italy) between June 2018 and March 2021 were retrospectively recruited for this single-center study. The following parameters were collected: age, results of DRE, Prostate-Specific Antigen (PSA) levels, and biopsy-based Gleason Score (GS). The retrospective use of data from clinical routine was performed according to institutional rules. The study was approved by the local Ethical Committee, ID number 5177. All procedures performed were in accordance with the ethical standards defined by the 1964 Helsinki Declaration and its later amendments.

Acquisition Protocol and Image Analysis
After biopsy-based confirmation of PC, all patients underwent 18 F-FECh PET/CT for staging purposes. Philips Gemini XL, Siemens Biograph mCT, and Siemens Biograph Vision V600 scanners were used. After a first non-contrast enhanced pelvic CT, 4MBq/kg of 18 F-FECh were intravenously administered to the patient in a supine position and immediately followed by a 5-min early PET acquisition of the pelvic region. A whole-body skull-base-tothigh PET/CT was performed about 60 min after 18 F-FECh administration. Images derived from Philips GEMINI XL were reconstructed by the LOR method, while images from both Siemens Biograph mCT and Siemens Biograph Vision V600 scanners were reconstructed by True-X TOF (ultraHD PET) iterative reconstruction algorithms (Table 4).
A Volume of interest (VOI) was manually drawn for each patient to contour the prostate gland showing suspicious 18 F-FECh uptake by applying a gradient-based threshold using a PET segmentation tool (LesionID, version 7.0.5 of MIM Encore Software Inc., Cleveland, OH, USA) and manual corrections were performed, when necessary, to avoid the inclusion in VOIs of spill-over areas due to 18 F-FECh accumulation in the bladder.
Conventional semiquantitative PET parameters were extracted from each VOI (t: tumor):

•
Maximum and mean Standardized Uptake Value, which indirectly estimates the maximum and mean values of 18F-FeCh concentration within the VOI by the normalization with the patient's body weight (tSUVmax and tSUVmean, respectively); • Metabolic Tumor Volume (tMTV), which represents the volume involving all the 18 F-FECh counts with at least 40% of SUVmax value; • Total Lesion Activity (tTLA as expression of tSUVmean × tMTV); • First-order radiomic features, such as tSkewness and tKurtosis, which describe the asimmetricity and the shape of distribution of 18 F-FECh values within the VOI, respectively. Additionally, a 2-cm spherical VOI was drawn on the spleen (Sp) to extract spleen SUVmean (SpSUVmean) and its standard deviation (SUVsd) as estimates of background intensity and noise, respectively.
Tumor-to-background ratio (TBR) and Signal-to-noise ratio (SNR) were calculated as follows: tSUVmax/SpSUVmean and tSUVmax/SUVsd, respectively. These ratios were introduced and included in the analysis to normalize the variability caused by the different scanners.
Advanced radiomic features were extracted from each prostatic VOI using the PET/CT module of an open-source IBSI-compliant platform (Moddicom) [21].

Statistical Analysis
Quantitative variables were described by mean and standard deviation, while qualitative variables by number and percentage.
A first analysis was performed considering the tumor's conventional PET parameters and first-order radiomic features (tSUVmax, tSUVmean, tMTV, tTLA, TBR, SNR, tSkewness, tKurtosis). For each binarized outcome (dichotomized PSA using median, dichotomized GS (6-7 vs. 8-9), and dichotomized DRE results suspicious vs. negative), logistic regression was performed to assess the predictive power of PET variables. Results were presented as odd ratios (OR) and corresponding 95% confidence intervals (CI). A multivariable model selection was performed among significant variables using Akaike's information criterion maximization.
The Mann-Whitney U test was used to evaluate the predictive role of advanced radiomic features with respect to the same binarized outcomes (PSA, DRE, GS). Radiomic bivariate models were built using the most relevant features (with p < 0.05 at univariate analysis) through logistic regression and trained/tested with a cross-fold validation test (80% vs. 20%, 10 repetitions). The best bivariate models were selected based on receiver operating characteristic (ROC) curves, mean area under the ROC curves (mAUC), and standard deviation (sdAUC).

Conclusions
Our study showed that tSUVmean only among semiquantitative parameters, tSkewness and tKurtosis as first order radiomic features were predictors of DRE results, while none of the semiquantitative parameters were able to be significantly correlated to the main clinical/histological parameters in PC patients, such as low/high PSA values and low and high-risk GS in PC patients referring 18 F-FECh PET/CT for primary staging. Larger prospective validation cohorts are recommended to explore if validated 18 F-FECh PET/CT radiomic models could represent an additional tool in the clinical management of newly diagnosed PC patients in the era of personalized medicine.

Conflicts of Interest:
The authors declare no conflict of interest.