This study aims to develop automated algorithms for prostate gland delineation and segmentation of prostate-specific membrane antigen (PSMA) expressing primary prostate cancer on PSMA PET/CT. Expert - generated prostate (n=315) and urinary bladder (n=87) contours were used for training a deep learning (DL) algorithm on low dose CT. PSMA PET/CT scans (n=86) from the randomized, multi-center proPSMA trial were used to design a watershed algorithm with region filtering (SUV cut-off 4) to delineate PSMA-avid foci within the DL-generated prostate gland contour. The algorithm, comprising DL and watershed, was subsequently used for evaluation in other PSMA PET/CT scans (n=170) from the proPSMA trial. The median Dice score coefficient and Hausdorff distance between the prostate cancer contours in the evaluation cohort delineated by the nuclear medicine physician (NMP) and our algorithm were 0.93 (IQR 0.89 to 0.96 - range 0 to 1) and 5.37 mm (IQR 4.07 to 11.88 mm - range 0 to 71.13 mm), respectively. The mean SUV max and mean SUV mean were 17.7 (SD 13.5, 95% CI 15.7-19.8) and 7.2 (SD 2.8, 95% CI 6.7-7.6) for the NMP and 17.8 (SD 13.8, 95% CI 15.7-19.9) and 6.9 (SD 2.8, 95% CI 6.5-7.3) for the algorithm, respectively. The mean volume intensity product (VIPPSMA = SUV mean x volume) and mean volume were 114 mL (SD 237.1, 95% CI 78.1-149.9) and 12 mL (SD 17.5, 95% CI 9.4-14.7) for the NMP and 123.8 mL (SD 262.8, 95% CI 84.1-163.6) and 13.6 mL (SD 21.2, 95% CI 10.4-16.8) for the algorithm, respectively. On per lesion analysis, 219 lesions were contoured by the NMP, compared with 245 by the algorithm (1.29, versus 1.44 per patient, respectively) with 2.9% false negative, and 11.8% false positive rate. Our algorithm can contour the primary prostate cancer on PSMA PET/CT in intermediate to high-risk prostate cancer with high-level agreement with human expert.
Alipour et al. (Wed,) studied this question.