Introduction
Current advances in prostate cancer (PCa) imaging, the transition towards precision medicine and tailored treatment, have led to the evolvement of focal therapy as treatment alternatives for localized PCa [
1]. Various treatment options like radiotherapy (RT), high-intensity focal ultrasound, or focal cryotherapy have been proposed to apply (i) a tumor-directed treatment by targeting solely the visible intraprostatic tumor volume while sparing the rest of the prostatic gland [
2] or by (ii) intensifying the treatment on the intraprostatic tumor volume while treating the rest of the prostatic gland conventionally [
3].
Multiparametric magnetic resonance imaging (mpMRI) is the standard of care (SOC) for the initial staging of patients with primary PCa [
1] and is momentarily utilized for targeted biopsy concepts [
4] and for focal therapy strategies [
5,
6] among other indications. However, mpMRI can underestimate the true tumor mass [
7‐
9]. The prostate-specific membrane antigen (PSMA) is selectively overexpressed in most PCa cells and can be traced by radiolabeled peptide ligands
68Ga-PSMA-11 in positron emission tomography (PET) [
10]. PSMA-PET/CT imaging has already been established as the gold standard for restaging in recurrent PCa after curative treatment [
11,
12]. Furthermore, there is growing evidence that PSMA-PET/CT is a suitable replacement for conventional imaging in the primary setting, providing superior accuracy at initial staging [
13,
14]. Considering intraprostatic tumor detection, previous studies suggested that PSMA-PET outperformed MRI in the detection of intraprostatic tumors with high sensitivities of 64–89% [
7,
15,
16]. However, PCa is renowned for its multifocality. Mouarviev and colleagues analyzed 947 prostatectomy specimens and only 22% of PCa were unifocal whereas 78% were multifocal with 2.24 lesions on average [
17]. For choline-PET, Souvatzoglou et al. [
18] proved that a significant amount of intraprostatic lesions might be missed by visual PET image interpretation due to small lesion size or lesion configuration. Currently, the characteristics of the visually undetectable lesions in PSMA-PET are unknown.
Obviously, the accurate detection of the entire clinically significant tumor mass is mandatory for a safe implementation of focal therapy strategies in patients with primary PCa. With the rise of big data analysis, the computer-based extraction of quantitative imaging parameters called radiomics features (RFs) may enable new concepts for personalized medicine. PET-derived RFs may be used as biomarkers to predict treatment outcomes and to characterize tumor biology non-invasively [
19]. For example, RFs from
68Ga-PSMA-11 PET images provided excellent results in Gleason score discrimination [
20].
The first aim of this study was to quantify and to characterize the clinically significant PCa lesions which have been missed by visual 68Ga-PSMA-11 PET image interpretation. Secondly, we analyzed whether basic clinical parameters and 68Ga-PSMA-11 PET-derived RFs can detect and characterize these visually undetectable intraprostatic lesions and their underlying multifocality. PCa distribution in co-registered whole-mount sections served as the reference in internal training and external validation cohorts.
Discussion
PSMA-PET/CT was introduced as an image modality with a high sensitivity (64–89%) for intraprostatic tumor detection [
15,
36] and the first studies are implementing PSMA-PET information to guide a focal RT [
37] by increasing the RT dose to the visible intraprostatic tumor mass. Most possibly, an ultra-focal treatment targeting solely the visible tumor mass in PSMA-PET images would reduce the treatment-related side effects due to better protection of the adjacent organs like rectum and bladder. However, treatment outcomes after ultra-focal therapies targeting solely the visible tumor mass in PSMA-PET images have not been reported yet. One main obstacle to this approach is that intraprostatic tumor lesions may be missed by visual PSMA-PET image interpretation due to two main reasons. First, detection of small lesions and of lesions with lenticular shape is hampered by the physical limitations of PET scanning systems as it was described for choline-PET previously [
18]. Second, intra- and intertumoral differences in PSMA expression were reported and approximately 10% of all intraprostatic lesions have no PSMA expression [
38]. Consequently, appropriate patient selection remains demanding for ultra-focal therapy approaches. To the best of our knowledge, this is the first report to examine the amount and the characteristics of the visually undetected intraprostatic tumor in PSMA-PET images. Additionally, we present two distinct RFs for their discrimination.
Taken together, 53% of the patients in our study had visually not detected PCa and 40% possessed clinically significant PCa with ISUP > 1. The visually undetected lesions were mainly small (median diameter 2.2 mm) and approximately one-third had a lenticular shape. This observation is in accordance with the study of Souvatzoglou et al. [
18]. The majority of the lesions (60.4%) were located in the peripheral zone which is also in concordance with previous studies [
17]. The invisible lesions were equally distributed in patients with differences in basic clinical parameters (like iPSA, pT stage, and NCCN risk group) and basic SUV parameters derived from GTV-PET (SUVmean and SUVmax). Consequently, other markers are warranted to predict the presence of invisible but clinically significant PCa lesion. Our group investigated RF from non-PCa-PET for the prediction of multifocality of PCa as a biomarker. Basic PET parameters from non-PCa-PET like SUVmean and SUVmax failed to discriminate patients with invisible lesions, although a strong trend for SUVmax was observed. However, two texture features (SAE and SZNUN) showed excellent performance, with ROC-AUC of ≥ 0.8–0.94 in the detection of invisible PCa lesions. The satisfying performance was repeatable in the external validation cohort with calculated sensitivities of ≥ 80% for non-resampled PET images. These findings may be explained by the technical properties of both RFs: quantification of the amount and distribution of small areas with PSMA uptake within the area of interest. The LBP filter was applied to enhance the detection of small areas with tracer uptake to notice even small PCa lesions. The LBP filter was already successfully applied by Kwak and colleagues [
39] in a computer-aided diagnosis system for the detection of prostate cancer, and on other malignant diseases as well [
34]. After resampling, the sensitivity decreased, suggesting that both RFs are very sensitive to image processing. Additionally, the spatial resolution in the validation cohort (2.3 × 2.3 × 2.7 mm) is lower than the median diameter of the missed lesions (2.2 mm) and subsampling the data to 2 × 2 × 2 mm will probably not help to reconstruct smaller lesions.
A reliable prediction of the presence of visually non-detectable but clinically significant PC might have impact on future therapeutic approaches and support decision-making. In patients with a high risk for clinically significant PCa in non-PCa-PET, ultra-focal therapy approaches targeting solely the visible tumor mass should be omitted and treatments targeting the entire prostatic gland (e.g., prostatectomy or conventional RT) should be offered. By applying predefined thresholds for both RFs, we calculated positive predictive values of > 0.85 in whole-gland analysis which indicates that the chance of falsely omitting focal treatments is low. Interestingly, we could reproduce the strong performance of both RFs from whole-gland analyses also by considering half of the prostatic gland in our training cohort. Prostatic halves with low risk for the presence of non-visible PCa lesions in PET may be spared from therapy (e.g., half-gland therapy) which may result in a significant reduction of treatment-related side effects [
5]. Future studies should also address the implementation of RF-based detection of PCa lesions in clinical routine workflows. First, an accurate segmentation of the visible tumor mass in PSMA-PET and mpMRI and the prostatic gland in CT or mpMRI should be performed manually or by using automatic tools (for example deep learning approaches [
40,
41]). Second, RF should be extracted by implementing already developed software tools for RF calculation [
42]. The entire workflow should be integrated into the standard software solutions for focal therapy or for targeted biopsy guidance. It should be mentioned that in 10–20% of the patients PCa lesions were missed by visual PET image interpretation and by implementation of RFs. The abovementioned lack of PSMA expression in approximately 10% of the PCa lesions may serve as an explanation for this result. In a study by Touijer et al. [
43], immunohistochemical analyses revealed that gastrin-releasing peptide receptor (GRPR) expression is not correlated with PSMA expression, suggesting that PET imaging targeting the GRPR may offer complementary information to PSMA-PET imaging. Thus, future studies should also assess the RF extraction from other imaging modalities in order to decrease the chance of missing lesions. It should be mentioned that our approach does not provide any information on the localization of the visually non-detected lesions. McGarry et al. [
44] implemented PCa radiomics to generate probability maps that generate a probability of malignancy in MR images on a voxel level. Whether this approach is applicable in PSMA-PET images should be addressed in further studies.
Sowalsky et al. [
45] postulated that a subset of PCa lesions with Gleason patterns 3 and 4 have a common origin. This finding supports a branched evolution model wherein Gleason pattern 3 and 4 tumors emerge from a common precursor. Additionally, Haffner et al. [
46] proposed that lesions with ISUP1 could be lethal for PCa patients as well. In contrary, recurrence-free survival after surgery or RT is significantly increased in patients with ISUP 1 [
47]. Consequently, other studies defined PCa lesions with ISUP > 1 as clinically relevant [
25]. Bravaccini et al. [
48] advocated that PSMA expression on PCa cells correlates with the ISUP score. However, in our study, both RFs failed to discriminate whether the lesions possessed clinically relevant PCa in terms of ISUP. This might be explained by the small size of the lesions in non-PCa-PET which probably results in minimal differences in measurable PET signal. Thus, differences in PSMA expression due to various ISUP scores will not result in significant diversity in PET signal and consequentially in RFs. Nevertheless, we suggest that both RFs deliver clinically relevant information despite the missing discrimination between ISUP 1 and ISUP > 1 lesions, since both RFs detected prostates with invisible PCa in visual PET analysis which revealed clinically significant PCa in the vast majority. Future studies integrating genomic information despite solely the Gleason score should re-analyze whether invisible but clinically significant PCa lesion with ISUP1 may be characterized with RF. A few papers on radiogenomics have been published already [
49,
50].
Due to an elaborate but labor-intensive PET histopathology registration protocol with high spatial resolution, a limitation of our study is the low number of patients in the training cohort. Considering the low number of patients and the non-significant effect of the standard clinical parameters (like PSA and NCCN risk group) on the distribution of the visually undetectable lesions, no additional multivariate regression analysis was performed in order to exclude confounding variables. In the validation cohort, only patients with PCa in non-PCa-PET were considered for RF analyses because of two reasons. Due to the relatively low resolution of histopathology information (5 mm between whole-mount slices) in the validation cohort, the chance of missing lesions in non-PCa-PET (approx. 1–5 mm diameter) was not negligible. Thus, it could not be excluded that false-negative findings of RF were due to the missed lesions in histopathology information instead of poor RF performance. Additionally, no dedicated co-registration protocol between histopathology slices and PET images was applied in this cohort. Likewise, a clear differentiation between small branches of the main tumor mass and autonomous lesions close to the main tumor mass was not always possible. In this study, the PSMA expression on the whole-mount tissue slices was not assessed by immunohistochemistry. Thus, it remains unclear how many of the visually undetectable lesions had no PSMA expression. Although MRI is the actual SOC for primary PCa staging, we decided to extract RFs from PSMA-PET images, as a higher sensitivity to detect primary PCa lesions has been reported for PSMA-PET when compared to MRI. The tracer that was injected in this study was
68Ga-PSMA-11. It is not clear whether the results of our study can be translated for
18F-labeled tracers. A recent study by Kuten et al. [
51] observed comparable results in PCa detection between
68Ga-PSMA-11 and
18F-PSMA-1007. However,
18F-PSMA-1007 seems to detect additional lesions of limited clinical relevance [
51]. The impact of PET-tracer on RFs is being controversially discussed [
52]. In this study, no multiple delineations of GTV-PET were obtained. Thus, it remains unclear how sensitive both RFs are to the quality of initial GTV-PET (and consequently non-PCa-PET) segmentation. However, in this work, we used a validated contouring approach for manual PET image segmentation and previous work revealed a very good interobserver agreement when using this technique [
23]. Additionally, both RFs performed well in whole- and half-gland analyses. This underlines the robustness of both RFs for invisible lesion discrimination independent of the analyzed volume of interest. Considering the limitations of our study, future work implementing histopathology reference with higher resolution should further validate the ability of RFs to detect primarily not visible but clinically significant PCa. We emphasize that RFs cut point values derived from our patient cohorts have been provided as a proof of concept and the optimal thresholds should be further investigated.
To conclude, visual 68Ga-PSMA-11 PET image interpretation missed small but clinically significant PCa in a relevant number of patients. However, two distinct RFs provided a very good performance in their detection. This may improve personalized diagnostic and therapeutic approaches for primary PCa by providing complementary information to visual PET interpretation.
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.