Session: MP61: Prostate Cancer: Localized: Surgical Therapy II
MP61-15: Skill-assessment during robot-assisted radioguided surgery - using artificial intelligence to extract kinematic metrics of DROP-IN gamma probe use
Urologist Leiden University Medical Centre, Leiden, Netherlands
Introduction: Radioguided prostate cancer (PCa) surgery has been greatly facilitated by the introduction of the steerable DROP-IN gamma probe technology. Reported indications include sentinel node (SN) resections and prostate-specific membrane antigen (PSMA)-targeted resections. Both make use of 99mTc-labeled imaging vectors. As result of the targeting mechanisms used the signal- and background intensities in these procedures, however, differ substantially. To study how this reflects on the surgical decision-making process we used computer-vision strategies to compare DROP-IN gamma probe use during SN- and PSMA-targeted surgeries. Methods: A DROP-IN gamma probe was applied during 45 robot-assisted PCa surgeries from 2018 to 2021 at single tertiary care referral centre: SN resections [n = 25; primary surgery; intraprostatic injection of indocyanine green (ICG)-99mTc-nanocolloid (~200MBq, 5-6h prior to surgery)], PSMA-targeted resections [n = 20; salvage surgery; intravenous injection of 99mTc-PSMA I&S (~550MBq, 16-22h prior to surgery)}. Preoperative imaging-roadmaps were created using SPECT/CT and PET/CT imaging. Using custom tracking algorithms, based on a neural network architecture, the DROP-IN probe trajectories were extracted from endoscopic videos. Trajectories were then processed to extract key kinematic performance metrics. To allow for skill-assessment, the metrics were incorporated in decision-making and dexterity scores. Results: The nodal signal intensities in PSMA- vs SN-targeted resections were 550 vs. 1200 counts in preoperative SPECT-CT scans (p=0.098) and 150 vs 1050 counts/s with regard to intraoperative probe readouts (p < 0.001). Due to the relatively high background signals observed during PSMA tracing a 2-fold reduction in the intraoperative signal to background ratio (SBR) (1.8 vs. 3.6; p<0.001) was observed in this setting. Kinematic analysis revealed significant differences in a number of metrics, e.g., target identification time (368 sec per PSMA lesion vs. 75 sec per SLN; p<0.001), number of probe pick-ups (5 vs. 1; p<0.001), and reduction in dexterity (281-fold; p<0.001) and decision-making scores (3.9-fold; p<0.001) during PSMA-targeted resections. Conclusions: DROP-IN technology facilitates robotic radioguided surgery in SN and PSMA-targeted resections. However, a profound performance difference was observed in favour of SN-targeted resections. In particular, the SBR <2 as found mostly in PSMA-targeted resections directly converted to a decline in dexterity and decision-making. These findings might be related to limited previously experience in PSMA-targeted surgery. SOURCE OF Funding: This research was supported by an NWO-TTW-VICI grant (TTW 16141) and with hardware by Eurorad S.A. and Intuitive Inc.
