TY - JOUR T1 - Automated Assessment of Prostatic PSMA Expression in SPECT/CT using Deep Convolutional Neural Networks - A Prospectively Planned Retrospective Analysis of Phase 3 Study MIP-1404-3301 JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 401 LP - 401 VL - 60 IS - supplement 1 AU - Karl Sjostrand AU - Lars Edenbrandt AU - Nancy Stambler AU - Adam Opanowski AU - Jens Richter AU - Konrad Gjertsson AU - Kerstin Johnsson AU - Vivien Wong AU - Jessica Jensen AU - Aseem Anand Y1 - 2019/05/01 UR - http://jnm.snmjournals.org/content/60/supplement_1/401.abstract N2 - 401Objectives: 99mTc MIP-1404 (1404) is a PSMA targeted imaging agent for the detection and staging of clinically significant prostate cancer. Manual assessment of tracer uptake in SPECT/CT images is time-consuming and subjective. Automated segmentation of organs and regions of interest and the subsequent quantification of PSMA expression may help physicians improve the accuracy and consistency of diagnosis. The study objective was to evaluate the performance characteristics of a prospectively locked deep learning algorithm (PSMA-AI) in the assessment of 1404. Methods: The study included 464 evaluable patients with very low, low, or intermediate risk prostate cancer, whose diagnostic biopsy indicated Gleason grade ≤3+4 and/or who were candidates for active surveillance (Study 1404-3301). All patients received an IV injection of 1404 and SPECT/CT imaging was performed 3-6 hours post-dose. They subsequently underwent either radical prostatectomy (low and intermediate risk) or prostate biopsy (very low risk). Clinically significant disease was declared in subjects with Gleason grade >3+4 or 3+4 with ≥10% pattern 4. PSMA expression was assessed by a target-to-background (TBR) value, defined by the ratio of the maximum uptake in the prostate and the average uptake in an adjacent background (muscle) region. Manual TBR was established by three independent, blinded readers using standard imaging workstations. PSMA-AI provided an automated TBR analysis attended to by three different independent readers. TBR values for all (3+3) readers and subjects were compared to the histopathological reference, yielding 6 receiver operating characteristic (ROC) curves. The area under ROC curve (AUC) was computed to determine the performance of the algorithm in distinguishing clinically significant from non-significant disease. The AUC of the three automated reads was compared to the AUC of the three manual reads. Further, inter-reader reproducibility was measured by correlation coefficients of log(TBR) for each pair of automated readers. PSMA-AI was developed and locked before any access to 1404-3301 data was granted. Results: The manual reads demonstrated AUCs of 0.62, 0.62 and 0.63. The reads with PSMA-AI demonstrated AUCs of 0.65, 0.66 and 0.66. The PSMA-AI performance in terms of AUC was higher than manual in all nine pairwise comparisons (3[asterisk]3=9), between the two reader groups, with statistically significant improvement observed in five cases (nominal p<0.05), not accounting for multiple comparisons. When measuring reproducibility, the log(TBR) correlation coefficients for pairs of PSMA-AI readers were 0.94, 0.97 and 0.98. The binary calls of a predefined threshold demonstrated a specificity of 95%, 96%, 96% and a sensitivity of 21%, 21%, 21% in detecting clinically significant disease in patients with very low, low, or intermediate risk of prostate cancer. Conclusions: The reads with PSMA-AI demonstrated an improvement over manual assessment in terms of speed, accuracy and repeatability. With improved performance characteristics, the assessment with PSMA-AI has the potential to augment clinical utility of PSMA imaging. This study has provided encouraging initial evidence in this direction. ER -