In this study, we have developed PSMA-PET-guided CT-based radiomic signatures for prediction of BRFS after sRT due to PCa recurrence using a multicenter cohort from three high volume centers. The developed radiomic signatures yielded good predictive performances and outperformed clinical signatures based on classical histological and clinical parameters. The radiomic model achieved significant patient stratification and demonstrated durable prediction of BRFS in time-dependent ROC analysis. To the best of our knowledge, this is the first study assessing CT-based radiomics in patients who underwent PSMA-PET based sRT and thus provides novel insights into this field of research.
Analysis of RFs have extensively been performed in primary prostate cancer patients (11). Most of these studies are based on MRI and demonstrated the ability of radiomics to non-invasively characterize and detect clinically significant PCa, extracapsular extension or predict BCR (26)(27)(28).
Fewer studies reported on CT-based radiomics and all of these were performed in the primary setting. Three studies developed CT-based radiomic classifiers with good performance to predict Gleason Score and risk groups (AUC 0.70–0.83) (29, 30)(31)
Based on PSMA-PET/CT scans, Peeken et al. developed a CT-based model to detect lymph nodes metastases, which outperformed conventional CT parameters with an AUC of 0.95 in external testing (32), addressing the limited ability of conventional imaging to detect PCa-positive lymph nodes. Acar et al. used CT-based RFs to differentiate between bone metastases and sclerotic areas with good accuracy (AUC 0.76) (33).
Since patients with BCR after surgery experience heterogeneous response rates (4, 34), our study aims to improve risk stratification with commonly available diagnostics for patients receiving sRT based on state-of-the-art diagnostics and to identify patients who might benefit from treatment de-intensification or intensification.
In our study, clinical signatures showed insufficient prognostic value for BRFS after sRT in the test sets, which reflects the deficiency of classical clinical and pathological parameters for prognostication demonstrated by retrospective and prospective studies. However, the developed radiomic signatures outperformed the clinical models with good prognostic values in the test sets. Radiomic signatures and particularly various feature selection methods outperformed clinical models, which demonstrates a certain robustness of these signatures. The inferior performance of the combined clinical- and radiomics signatures might be explainable due to the poor prognostic value of clinical parameters and the low patient number for effective model building.
Since no other studies evaluated CT-based radiomics to predict BCR, we cannot directly compare our signatures with other CT-based models. Nevertheless, in comparison with mpMRI derived RF the radiomic models in our study performed similarly well with a C-index of > 0.7, considering different clinical scenarios between these studies. DCA demonstrates a net benefit of the radiomic signatures, suggesting that clinical utilization of radiomics can help to identify patients who are at higher risk of BCR after sRT. Whether these patients benefit from intensified treatments and if, which kind of treatment intensification is optimal, needs to be evaluated in future studies.
CT-based radiomics might in future play an even more important role, since technical advantages such as dual-energy CTs provide more image information and may thus allow for more differentiated radiomic analyses. Additional, the prognostic capability of PSMA-based radiomic signatures need to be evaluated in future studies.
Due to the small patient number, we were not able to separate an external testing cohort, but rather obtained high statistical robustness by applying a nested cross validation approach. Future studies should focus on external validation to demonstrate transferability of models.
The mean intensity within the VOI was selected as the most important RF. Lower intensity values were associated with decreased BRFS. To provide a simple cut-off metric, we applied the maximally selected rank statistics. A cut-off of 19.7 HU was determined as optimal cut-off point for BCR. However, unlimited reduction of HU values is not plausible, since we expect local recurrence to have HU values greater than fat tissue. Thus, this cut-off value should be validated in further studies. Moreover, we provide a univariate Cox model and nomogram trained on the complete cohort for future external validation.
There are several limitations in our study. First we want to mention the retrospective character ant possible selection bias. Secondly, we have included patients with LR and NR in this analysis, who experience different outcomes. Separation of both cohorts would have resulted in an insufficient sample size. Nevertheless, development of CT-based radiomics signatures might be influenced to a lesser extent through this heterogeneity in comparison to functional imaging methods. Thirdly, we used an internal validation due to the low number of patients within each institution. However, we applied a sophisticated nested cross-validation approach to overcome methodological disadvantages. The inclusion of patients that received ADT may have biased optimal outcome predictions. Again, exclusion of these patients would have significantly reduced the sample number. Finally, the FU in our cohorts is relatively short with a median FU of 29 months.