Prospectively planned and Independent Validation of aPROMISE in a Phase III CONDOR Study for Rapid Lesion Detection and Standardized Quantitative Evaluation for 18F-DCFPyL (PSMA) Imaging in Prostate Cancer.

Introduction: Credentialed quantification of the whole-body tumor burden by PSMA PET in patients with prostate cancer is required to enable PSMA PET as an imaging biomarker. aPROMISE is a deep learning automated platform quantitatively assess PSMA PET images. The objective of this study was to evaluate aPROMISE against manual reads by blinded independent readers.

Methods: 208 patients with biochemical recurrent prostate cancer enrolled in the PyL-3301 (CONDOR) study (NCT03739684) were included in the analysis. The PSMA PET/CT scans were analyzed by three independent readers (R1, R2, R3) with and without – automated Prostate Cancer Molecular Imaging Standardized Evaluation (aPROMISE V2.0). Detection and quantification of PSMA-positive lesions by aPROMISE is performed by the use of indices (automated PSMA score [aPSMA-score]) in the prostate fossa, pelvic lymph nodes or extra-pelvic lesions following the miTNM classification. The lesion aPSMA-score <m:ctrlpr></m:ctrlpr> I T for tissue type/region T provides a quantitative assessment of disease burden based on reference organ uptake and lesion volume: <m:ctrlpr></m:ctrlpr> I T = <m:ctrlpr></m:ctrlpr> i∈T <m:ctrlpr></m:ctrlpr> s i <m:ctrlpr></m:ctrlpr> v i where <m:ctrlpr></m:ctrlpr> s i is the SUV-mean for lesion i normalized by reference organ and <m:ctrlpr></m:ctrlpr> v i is the corresponding lesion volume. The aPROMISE reads were compared to the Standard of Truth (SOT) data to distinguish true positive lesions from false negatives. The SOT (consensus read from two independent readers) comprises histopathology and/or conventional imaging. Fleiss Kappa (k) agreement and Intraclass correlation coefficient were used to compare lesion staging and lesion SUV reproducibility, respectively, of readers with and without aPROMISE. The success criteria were predefined by the sensitivity and ICC of aPROMISE as good or better compared to manual reads. For the endpoint to be declared a success, the test must succeed in at least 2 out of the 3 readers.

Results: All 208 patients were assessed by aPROMISE. A total of 323 lesions were identified by SOT. The Fleiss kappa agreements among the three readers in staging assignment with aPROMISE (miT 0.70; miN 0.71; miM 0.62) and without aPROMISE (miT 0.70; miN 0.71; miM 0.62), were in the 0.6-0.8 interval. In the sensitivity analysis against SOT, one reader demonstrated significant improvement of lesion detection with aPROMISE (R1-70%, 95% CI:65-75%) compared to without (62%, 95%CI: 56-67%), p=0.006. In the other two readers, there was moderate improvement, but no significant difference with aPROMISE (R2-78%, 95%CI:74-83%; and R3-77%, 95%CI:72-81%) compared to without (R2-77%, 95%CI:72-82%; R3-76%, 95%CI:71-81%), p=0.760 and p=1.00, respectively. Median reading times for R1, R2 and R3 were, 1.4 min, 1.2 min, and 1.4 minutes, respectively. The interquartile range of all reading times combined was 2.1 minutes. Read durations were not measured in the manual CONDOR study, however average reading time was reported to be approximately 15 minutes. The reproducibility of per-lesion SUV measurements among the 3 readers was significantly higher with aPROMISE (ICC 0.98 95% CI 0.96 to 0.99) than without aPROMISE (ICC 0.90 95% CI 0.88 to 0.91) p<0.0001.

Conclusions: The study demonstrated the higher efficiency and consistency of the aPROMISE platform while maintaining the diagnostic accuracy of the PSMA imaging. The quantitative aPSMA-score warrants future clinical investigation to define its clinical context of use as an imaging biomarker.