Abstract
241742
Introduction: PYLARIFY-AI (AI) is an application enabled with deep learning for standardizing the interpretation and reporting of PSMA PET. CONDOR demonstrated its ability for rapid lesion detection and reproducible quantitative assessment, while maintaining diagnostic accuracy for PSMA PET imaging. We have encountered the detection of a significant number of probably FP or equivocal (E) foci of activity with this application. OBJECTIVE: Identify potential factors contributing to the occurrence of possible FP detected by AI.
Methods: 10 patients with progressing prostate cancer underwent PYLARIFY PET/CT. Interpretation performed by experienced nuclear medicine physician, followed by AI analysis. Prostate/seminal vesicles (P), node (N), bone (B) and organ mets (C) were documented for interpreter and AI. Reported foci classified as TP if probably true or equivocal. FP for very unlikely to represent mets. Correlation analysis performed comparing FP, BMI, Camera (newer TOF or older BGO), tracer dose and PSA level.
Results: 9/10 patients had PSMA uptake in prostate bed reported by interpreter and AI. AI detected pixel focus between seminal vesicles (SV), probably noise.
8 PSMA avid nodes detected by interpreter measured <4 mm, AI detecting 3/8 of these. 1 E node detected by AI and 3 other avid nodes detected by AI but not originally by interpreter. 44 FP nodes detected by AI, most of these projected on bowel and ureters. 3 foci along the bladder activity border and 1 presacral focus likely inflammatory. AI detected bone activity in the seminal vesicle (pixel), muscle below right ischium, vessel and right hand.
3 foci detected by interpreter and AI. 3 areas of diffuse mild E uptake associated 3 different vertebral fractures, 1 detected by AI. Very mild focal uptake in the skull detected by AI but not seen by reader, correlating with a 7 mm expansile lucency. 61 probably FP bone foci detected by AI, most correlating with bone cortex and facet joints. Other small and pixel sized foci in ribs without clear CT correlate and within noise range. Some foci in the skull correlate with lacrimal gland (3), scalp (1) and transverse sinus (1) physiologic activity. 1 focus correlating with vascular calcification and 2 foci in the same patients correlating with bowel activity.
Large highly PSMA avid splenic mass seen by interpreter but not AI. AI detected E mild uptake in the right hilum of same patient, likely a reactive node. 3 probably FP foci with 2 areas of mild diffuse uptake correlating with ground glass opacity in 1 patient, probably reactive or inflammatory. 1 focal area associated with misregistration of liver activity in the right lower lung zone with no lung abnormality on CT.
Strong positive correlation between FP and BMI (0.72). Strong positive correlation between FP and TOF vs BGO but likely bias since patients with higher BMI were scanned with TOF camera. No strong correlation of FP and dose (7.9-10.8mCi).
AI was not able to detect a large highly PSMA avid splenic lesion. AI was able to detect 4 PSMA avid nodes not originally seen by interpreter. 5 highly PSMA avid tiny nodes not detected by AI. 1 tiny PSMA avid expansile lesion in skull not detected by interpreter.
Conclusions: As previously documented, we also found a significant number of FP which may interfere with image analysis. We anticipate that knowing the contributing factors may help the AI technique improve reducing them. A strong positive association between significant FP metastatic foci detection by PYLARIFY-AI and BMI seen, which is unclear if it is only attributed to SNR degradation or if CT organ segmentation and/or blob detection techniques may have limitations in its implementation with higher BMI. No clear association between FP with dose or camera type needs further evaluation.
AI was not able to detect some < 4 mm avid N, maybe due to limitations in CT organ segmentation. Large splenic avid lesion not detected by AI, probably due to limitation of blob detection technique.
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.