%0 Journal Article %A Saeed Ashrafinia %A Mohammad Salehi Sadaghiani %A Pejman Dalaie %A Rima Tulbah %A Yafu Yin %A Kevin Leung %A Ryan VanDenBerg %A Jeffrey Leal %A Michael Gorin %A Martin Pomper %A Arman Rahmim %A Steven Rowe %T Characterization of Segmented 18F-DCFPyL PET/CT Lesions in the Context of PSMA-RADS Structured Reporting %D 2019 %J Journal of Nuclear Medicine %P 1565-1565 %V 60 %N supplement 1 %X 1565Objectives: A prostate-specific membrane antigen (PSMA) reporting and data system (PSMA-RADS version 1.0) was recently introduced for PSMA-targeted PET imaging studies, in order to classify individual findings into categories that reflect the likelihood of the presence of prostate cancer (PCa). This structured reporting framework is based on imaging findings including anatomic location and intensity of radiotracer uptake. We aim to report on imaging features and segmentation characteristics of 18F-DCFPyL (PyL) lesions as categorized based on the PSMA-RADS scoring scheme. Our work also provides more insight into segmentation of PyL PET images in general. Methods: The dataset contains 277 PCa patients imaged with PyL PET/CT at 60 min post-injection. Four trained nuclear medicine physicians manually segmented 3,794 lesions regions of interest (ROI), in addition to a spherical ROI (30 mm diameter) in the liver for reference, yielding a total of 4,071 ROIs. They subsequently performed PSMA-RADS scoring for every lesion utilizing both PET and CT images. Detailed underlying anatomic location was also recorded (572 sites), categorized into 9 broad anatomic categories. SUVmax, tumor volume and effective thresholding level for each manually segmented ROI was also calculated. The effective threshold was computed as min{ SUVminROI, (SUVminROI+SUVmaxring)/2} (expressed as percentage of SUVmaxROI), where SUVmaxring was calculated outside of the ROI on the immediately adjacent 1 to 2 rings of voxels. Results: The distribution of lesions was as follows: 280 PSMA-RADS 1As, 632 1Bs, 830 2s, 337 3As, 147 3Bs, 30 3Cs, 40 3Ds, 615 4s, and 838 5s, showing higher prevalence of intermediate-to-high uptake lesions with scores 4 and 5, followed by lower uptake lesions such as 3A. Axillary and inguinal lymph nodes together accounted for 599 lesions, and were generally considered PSMA-RADS-2 as they were not a common site of PCa. While the internal ROI segmentation thresholds varied highly within each PSMA-RADS score, the thresholding was less variable within the same location. Generally, lesions with intermediate uptake in 1B, 2, 3A, 3B, 3C and 3D had low SUVmax (average 2.2, 1.7, 2.8, 1.8, 6.6, and 2.1, for each score, respectively) compared to 4 and 5 (average SUVmax of 9.8 and 14.1, respectively). Linked to this, the former group of lesions had a higher effective segmentation threshold compared to the latter, as their uptake was closer to the background. In nearly half the ROIs, the immediate neighborhoods contained voxel(s) with value >SUVminROI indicating that more sophisticated thresholding techniques are required for automatic/semiautomatic segmentation of PyL PET images. This was observed even for scores 4 and 5 with high uptakes. Proximity to normal areas with intermediate uptake (vessels, bowel loops) for scores 1B, 2, 3A, and 3B, or high uptake (kidney, bladder) for 4 and 5 contributed to that effect. PSMA-RADS scores for lesions with highest volumes (mean/median in mL) were 1B (14.2/3.1), 3C (16.4/13.4), and 5(5.6/2.1). The median volume across all lesions was 1.14 mL with only 7.6% of the lesions larger than 10 mL. Conclusions: Median effective segmentation thresholds were low (43%, 35%) for PSMA-RADS scores 4 and 5 but higher (61.4% ~ 70%) across 2 and all 3 subcategories. The challenging segmentation task for PyL PET images requires supervision by an expert or use of an advanced method such as deep learning. The relatively small sizes of the PyL lesions demonstrate its high focal concentration and specificity; yet, it also indicates that a radiomics analysis of these lesions may not provide new, significant information. Nonetheless, the differences in parameters related to each of the PSMA-RADS categories suggests that a sufficiently large dataset of lesions could be used with artificial intelligence applications to automate lesion categorization. View this table:Distribution of Lesion Underlying Tissue Sites for Each PSMA-RADS Score %U