Deep learning algorithm improves identification of men with low-risk prostate cancer using PSMA-targeted ^99mTc-MIP-1404 SPECT/CT.

e16572 Background: Previous work has shown that the degree of expression of prostate-specific membrane antigen (PSMA) correlates with prostate cancer (PCa) grade and stage. We evaluated the additive value of a deep learning algorithm (PSMA-AI) of a PSMA-targeted small molecule SPECT/CT imaging agent (99mTc-MIP-1404) to identify men with low risk PCa who are potential active surveillance candidates. Methods: A secondary analysis of a phase III trial (NCT02615067) of men with PCa who underwent 99mTc-MIP-1404 SPECT/CT was conducted. Patients with a biopsy Gleason score (GS) of ≤6, clinical stage ≤T2, and prostate specific antigen (PSA) < 10 ng/mL who underwent radical prostatectomy (RP) following SPECT/CT were included in the present analysis. SPECT/CT images were retrospectively analyzed by PSMA-AI, which was developed and locked prior to analysis. PSMA-AI calculated the uptake of 99mTc-MIP-1404 against the background reference (TBR). The automated TBR of 14 was used as a threshold for PSMA-AI calls of positive disease. Multivariable logistic regression analysis was used to develop a base model for identifying men with occult GS ≥7 PCa in the RP specimen. This model included PSA density, % positive biopsy cores, and clinical stage. The diagnostic performance of this model was then compared to a second model that incorporated PSMA-AI calls. Results: In total, 87 patients enrolled in the original trial contributed to the analysis. The base model indicated that PSA density and % positive cores were significantly associated with occult GS ≥7 PCa (p < 0.05), but clinical stage was not (p = 0.23). The predictive ability of the model resulted in an area under the curve (AUC) of 0.73. Upon adding PSMA-AI calls, the AUC increased to 0.77. PSMA-AI calls (p = 0.045), pre-surgery PSA density (0.019) and % positive core (p < 0.004) remained statistically significant. PSMA-AI calls increased the positive predictive value from 70% to 77% and the negative predictive value from 57% to 74%. Conclusions: The addition of PSMA-AI calls demonstrated a significant improvement over known predictors for identifying men with occult GS ≥7 PCa, who are inappropriate candidates for active surveillance. Clinical trial information: NCT02615067.