Prostate cancer (PCa) ranks the second highest globally in terms of the prevalence of male malignancies, with more than 1.4 million new cases diagnosed in 2020 (1). High-risk PCa accounts for over one-third of the newly diagnosed PCa population, with a three-fold greater risk of developing distant metastasis compared to their low-risk counterparts. The management strategies of these two cohorts differ drastically. The low-risk PCa usually requires only active surveillance, while high-risk ones require combined modality therapy such as surgery, radiotherapy, systemic chemotherapy or hormonal therapy (2). Optimising treatment strategy for the high-risk is challenging.

Clinically undetectable occult pelvic nodal metastasis is commonly present in up to 40% of high-risk PCa patients (3). However, reliable detection of occult pelvic lymph node (PLN) metastasis is yet available for clinical use (4). It remains as an unresolved clinical challenge as to whether PLN should be prophylactically treated. The survival benefits and toxicity profiles of prostate-only radiotherapy (PORT), or prophylactic whole-pelvic radiotherapy (WPRT) were vigorously investigated in large randomised controlled trials (e.g. RTOG-9413), national database analysis and retrospective studies among the western population (5–9). Hence, the trade-off between better survival with WPRT and reduced toxicities with PORT is still highly debated.

Although the 5-year survival of high-risk PCa patients drastically increased by 23% over the years with greater emphasis on health-related quality of life (HRQoL) (10, 11), over 50% of high-risk PCa patients receiving PORT experienced recurrence, which was far higher than the 12.5% recurrence from the WPRT cohort (6). It is evident that a more refined subgrouping is necessary to predict which high-risk patient receiving PORT would experience recurrence within 5 years to support the clinical decision. Two commonly used conventional risk stratification tools are the National Institute for Health and Care Excellence (NICE) guideline and the Roach formula. The NICE guideline stratified patients into low, intermediate and high risk by clinical (c)T stage, prostate-specific antigen (PSA) serum concentration and Gleason score (GS) (12). Although all are prognostic markers (13–15), the cT stage is determined by digital rectal examination (DRE) that is subjected to high interobserver variability because (16) only a small portion of the prostate is palpable (17). The Roach score (RS) is also commonly used for risk stratification, based on the PSA and GS (18). A score of ≥15% is considered high-risk. A recent study reported that the RS has statistically significantly higher predictive power than the NICE guideline, with a concordance index of 0.724 and 0.715 respectively (13). However, the RS tends to overestimate the risk of occult PLN disease by 2.5 to 4 times among high-risk patients. This would result in the over-treatment of patients and compromise the therapeutic index (19, 20). Therefore, both NICE and RS may not be sufficiently effective in the era of precision medicine. There is a growing demand for developing a more refined, personalised risk stratification method for predicting treatment outcomes of high-risk PCa patients.

Recent advancement in artificial intelligence and radiomics accelerates the development of precision and personalised medicine. Radiomics adopts high-throughput methods to extract high-dimensional radiological features, transforming them into imaging biomarkers to improve clinical decisions (21). Magnetic resonance imaging (MRI) has been extensively studied, employing derived radiomic features for diagnosis or risk prediction of PCa. They showed promising classification performance on clinical endpoints such as GS or biochemical recurrence (BCR) (22–26). By contrast, very few investigations were conducted on the prognostic value of imaging biomarkers derived from the pCT of high-risk PCa patients, despite that the prognostic power of CT-based radiomics has been widely reported in other types of primary solid tumours such as non-small cell lung cancer, nasopharyngeal carcinoma and renal cell carcinoma (27–30). Franzese et al. (31) was the only study employing pCT-derived radiomic features to predict metastasis-free survival in PCa patients treated by external beam radiotherapy (EBRT). However, the included patients in their study were treated with various kinds of treatment, including trans-urethral resection of the prostate, high-intensity focused ultrasound, and EBRT. The prognostic power of their model specifically on PORT-treated high-risk PCa patients remains to be explored.

To our best knowledge, this is the first study to investigate the feasibility of utilising pCT-derived radiomic features and clinical attributes to predict 5-year progression-free survival (PFS) in high-risk localised PCa patients following PORT. Recently, a systematic review in prostate radiomics suggested that incorporating clinical features into a radiomic model may improve its clinical utility (32). Given the inherently heterogeneous nature of the disease, conventional clinical factors may provide additional prognostic value (33). The success of this study may provide insightful findings for clinicians to optimise treatment strategies for managing high-risk PCa patients.

For the first time, we demonstrated that combining CT-based radiomic (unfiltered-shape-flatness, LoG-1mm-filtered RMS) and clinical attributes (PSA, GS group, RS group and GTV volume) provided superior prognostic value for 5-year PFS in high-risk PCa patients following PORT. The DeLong test revealed that the ROC-AUC of the RC combined model was significantly higher than those of the R and C models in the independent validation cohort. The RC model had the highest overall accuracy and YI. Only the RC model score significantly classified patients into “progression” (PFS-1) and “progression-free” (PFS-0) groups according to their 5-year PFS in both training and independent validation cohorts (both p< 0.05). These findings suggested the potential of the RC combined model in supporting clinicians to implement personalized treatment for this vulnerable patient subgroup in the future. For instance, if a patient is classified into the “progression” (PFS-1) group ahead of the commencement of the PORT treatment, clinicians may consider a more aggressive therapy (e.g., WPRT) for improving the prognosis of the given patient.

The two identified radiomic features (unfiltered-shape-flatness and LoG-1mm-filtered RMS) are in line with previous studies using CT images (27, 28, 48, 49). Unfiltered-shape-flatness is a shape feature calculated by the square root of the least axis length divided by the major axis length (48). A value closer to zero indicates the tumour is flatter. This feature was reported in another CT radiomic model for predicting tumour response in lung cancer patients receiving EBRT (49). Additionally, the RC-score was negatively correlated to the flatness, with a lower flatness value indicating a poorer prognosis. Such a finding is consistent with Khodabakhshi’s study, which predicted the OS of patients with renal cell carcinoma (27). The feature LoG-1 mm-filtered RMS is calculated by the mean of all squared intensity values in LoG-1mm-filtered ROI (48). Similar to flatness, it is negatively correlated to the RC-score. The RMS has been reported in a CT radiomic model study for lung adenocarcinoma, in which a lower value was associated with poorer PFS and OS (28). The prognostic implications of these features have not been reported for PCa, and the association between the selected pCT radiomic features and biological properties remains unclear. Indeed, the RC model was dominated by the R features since the absolute SCC of R features with RC-score was ~2.6 times higher than those of clinical features (Supplementary Material). Further investigation is needed to explore the biological mechanism of radiomic features.

Another notable finding is the lack of textural features in the RC combined model. It may be attributed to the intrinsic property of pCT images and the clinical endpoint of our study. Most prostate radiomic studies use MRI-derived features for prognostic prediction (22–24, 50). In a study using T2-weighted MRI, features with the highest predictive value originated from the gray-level run-length matrix (GLRLM) texture feature class (23). Another two studies using apparent diffusion coefficient (ADC) MRI for radiomics modelling (23, 24) suggested that the gray-level co-occurrence matrix (GLCM) texture feature class contained the most predictive features. However, in a prostate radiomic study (51) exploring interfraction cone-beam CT, both shape and first-order features have excellent capability in predicting patient outcomes, which was similar to our study. In addition, these two features were also found to be capable of predicting PFS in both nasopharyngeal carcinoma (NPC) and NSCLC (29, 30). Meanwhile, texture features dominate in MRI-based NPC radiomic studies (52, 53). These observations suggested that the feature class selected for PFS prediction could be influenced by the choice of imaging modality. This may be explained by the inherently lower soft-tissue contrast characteristic in CT than in MRI, resulting in less prominent texture features that may be relevant to the PFS. Texture features in CT and MRI have been regarded as the manifestation of tumour heterogeneity. Hence, the dominant feature class may vary according to different clinical endpoints. For instance, texture features are often selected in both CT and MRI prostate radiomic studies when GS is chosen as the clinical endpoint (25, 42). It is not surprising since GS is the gold standard for characterising prostate heterogeneity, while texture features measure the same physical property. The current study demonstrated that both shape (unfiltered-shape-flatness) and first-order features (LoG-1mm-filtered RMS) have similar predictive performance compared to MRI radiomic studies (23, 24, 50). Nonetheless, a further study involving a larger external validation cohort is needed to validate the performance of these two feature classes in predicting the PFS of high-risk PCa patients.

The C model constructed based on clinical factors had the poorest performance. Univariate analysis did not show any statistical significance of these clinical features between PFS-1 and PFS-0 patients in both cohorts. The most probable explanation is that these clinical features are homogeneous within the high-risk subgroup of PCa patients (32). Thus, using the C model alone would lead to the poorest prediction performance. Statistically, the C model also failed to differentiate the 5-year PFS status in both cohorts. These findings are in line with those reported by Fernandes et al. (23) and Bourbonne et al. (26), who have also incorporated PSA and GS in their clinical models. They also retrospectively investigated the clinical predictive model for high-risk PCa patients treated by EBRT (23) and radical prostatectomy during a 5-year follow-up (26). The sensitivity and specificity of the C model were also comparable to those reported by Bourbonne et al. (26) (Sensitivity: 0.53 vs 0.68; Specificity: 0.53 vs. 0.59). These clinical features, however, are not without predictive values when combined with the R model. It is demonstrated by the increase in overall accuracy in both the training and independent validation cohorts after combining R with C models. The potential complementary role between R and C features for prognosis warrants future investigation (54).

Overall, this study explored the potential of combining pCT-derived radiomic and clinical features in the prognostic prediction of high-risk PCa patients receiving PORT. Our study has demonstrated that the classification performance of the combined RC model was comparable to the combined MRI-derived radiomic and clinical models. Among all included features, the shape and first-order features are considered more intuitive than other complex features when interpreting the classifier in predicting 5-year PFS of high-risk patients in the clinical setting (51). Our patient stratification methodology is highly standardised by adopting a temporal validation as in type-2b study according to the TRIPOD guideline (39). Moreover, the use of pCT is preferable due to its higher standardisation, repeatability and calibration of CT over diagnostic MRI (33). For instance, HU in pCT directly quantifies the electron density of the tissue while the pixel value in MR is arbitrarily allocated.

This study has several limitations. First, the sample size is relatively small. This can be attributed to the strict inclusion and exclusion criteria enforced to ensure high-quality radiomics and clinical data. Nonetheless, numerous studies have also contributed insightful findings to the community with a similar sample size as our work (32, 55–57). Second, this work only performed modelling based on data from a single centre as COVID-19 has restricted the research team from conducting data collection from multiple centres. Third, owing to the small sample size and temporal stratification, a class imbalance exists in which only 10% are PFS-1 patients. Although a similar situation was observed in Bosetti’s (2020) work analysing cone-beam CT performance for predicting prostate cancer clinical progression, further study with a larger sample size and proportion of PFS-1 patients would be beneficial to minimise modelling and evaluation bias. Minority class-boosting techniques should also be implemented if appropriate. Fourth, an external testing set was not employed. This would be essential to demonstrate satisfying model generalisability in a multi-centre study before actual clinical application (58). Moreover, Ridge regression, which was a relatively straightforward modelling strategy, was adopted in this work for demonstrating the proposed feasibility of using radiomic-clincial factors to predict patient outcomes. To render the model fit for clinical application, more robust modelling methodologies such as non-linear machine learning techniques and random oversampling should be incorporated when processing multi-centre data sets.

This study demonstrated the feasibility and potential of using pCT-derived radiomic and clinical features for predicting 5-year PFS in high-risk PCa patients receiving PORT, which is an important clinical research gap that previously lacks investigation The RC combined model provided statistically superior predictive performance than both R and C models in the independent temporal validation cohort. These findings lay the ground for the future development of a combined radiomic-clinical model involving robust modelling techniques, multicentre data and external validation. We hope that this work will bring attention to the academic community and encourage future work to address this on a larger scale towards clinical implementation. Ultimately, it could potentially act as a supportive decision tool predicting the outcome of different treatment regimens to facilitate personalised management of high-risk PCa patients.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by Human Subjects Ethics Sub-committee of The Hong Kong Polytechnic University and Kowloon West Cluster Research Ethics Committee of the Hong Kong Hospital Authority. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

JCC, SL, WSL and SWL contributed to the study design, methodology development, results interpretation, and manuscript review. WL offered administrative and material support for data collection. JCC, AOL, and JCK collected patients’ clinical and imaging data. JCC, SL and AOL contributed to image pre-processing and feature extraction. JCC, SL, CCL, AOL and JCK constructed the models. JCC wrote the manuscript. WSL co-supervised the study. SWL supervised the study. All authors contributed to the article and approved the submitted version.