A Prognostic Risk Score for Prostate Cancer Based on PSMA PET-derived Organ-specific Tumor Volumes

Abstract

Background Prostate-specific membrane antigen (PSMA) PET has high specificity in localizing primary tumors and metastases in patients with prostate cancer, but the individual overall survival probability is still difficult to estimate. Purpose To develop a prognostic risk score using PSMA PET-derived organ-specific total tumor volumes for predicting overall survival in patients with prostate cancer. Materials and Methods Men with prostate cancer who underwent PSMA PET/CT from January 2014 to December 2018 were evaluated retrospectively. All patients from center A were split into training (80%) and internal validation (20%) cohorts. Randomly selected patients from center B were used for external validation. Organ-specific tumor volumes were automatically quantified from PSMA PET scans by a neural network. A prognostic score was selected using multivariable Cox regression guided by the Akaike information criterion (AIC). The final prognostic risk score fitted on the training set was applied to both validation cohorts. Results A total of 1348 men (mean age, 70 years ± 8 [SD]) were included, with 918 patients in the training cohort, 230 in the internal validation cohort, and 200 in the external validation cohort. The median follow-up time was 55.7 months (IQR, 46.7-65.1 months; >4 years; 429 deaths occurred). A body weight-adjusted prognostic risk score integrating total, bone, and visceral tumor volumes obtained high C index values in the internal (0.82) and external (0.74) validation cohorts, as well as in patients with castration-resistant (0.75) and hormone-sensitive (0.68) disease. The fit of the statistical model for the prognostic score was improved compared with a model containing total tumor volume only (AIC, 3324 vs 3351; likelihood ratio test, P < .001). Calibration plots ascertained good model fit. Conclusion The newly developed risk score that included prostate-specific membrane antigen PET-derived organ-specific tumor volumes had good model fit for predicting overall survival in both internal and external validation cohorts. Published under a CC BY 4.0 license. Supplemental material is available for this article. See also the editorial by Civelek in this issue.

Figure 1 –. Patient selection flowcharts.

Patients were recruited at the University Hospital Münster and the University Hospital Vienna. PSMA = prostate-specific membrane antigen

Figure 2 –. Overall survival prognostication in the training cohort.

Patients of the training cohort (n = 918) were stratified into those with no tumor (zero tumor volume) and those with detectable tumor; the latter were additionally stratified according to the quartiles of tumor volume (Q1-Q4). Overall survival was plotted by Kaplan-Meier estimation. The total tumor volume, distant lymph node tumor volume (M1a), bone tumor volume (M1b) and soft tissue tumor volume (M1c) are plotted separately. Additionally, a subpopulation treatment effect pattern plot (STEPP) analysis for the 60-month follow-up timepoint is shown for each prognostic marker.

Figure 3 –. Quality of the prognostic risk score in the training cohort.

(A) Patients with a risk score greater than the median of the total cohort were grouped by quintiles (Q1-Q5). (B) For the total training cohort, a subpopulation treatment effect pattern plot (STEPP) analysis is shown for the 60-month timepoint. The STEPP-analysis depicts a decreasing survival probability with an increasing risk score. (C) For the total training cohort, time dependent area under the receiver operating characteristic curve (AUC) values with 95% CI of the multivariable risk score prognosticating overall survival are shown for each follow-up timepoint. (D) The calibration plot shows the overall survival prediction for the total training cohort at the 60-month timepoint.

Figure 4 –. illustrative PSMA-PET cases of patients with differing outcomes and tumor volumes.

The PSMA-PET-based maximum intensity projections (MIPs) of 4 exemplary patients are shown. Additionally, automatically segmented tumors or metastases are presented as orange overlay. The prognostic risk score was used to calculate the expected 24-months overall survival probability and its corresponding 95%CI.

Figure 5 –. Internal and external validation of Cox model

(A) The time dependent area under the receiver operating characteristic curve (AUC) values with 95% CI for the overall survival prognostication by the prognostic risk score in the internal validation cohort are shown for each follow-up timepoint. (B) The prognostic risk score was used to group the patients in the external validation cohort by the median (yellow > median, blue ≤ median); patients with high linear predictor values have shorter overall survival (63.4 months vs. median not reached, p<.05). (C) The time dependent AUC values with 95% CI for overall survival prognostication in the external validation cohort are shown. (D) The calibration plot depicts the survival prediction by the prognostic risk score at the 60-month timepoint in the external validation cohort.