UB-DARTS

Surgical artificial intelligence (AI) has the potential to improve patient safety and clinical outcomes. To date, training such AI models to identify tissue anatomy requires annotations by expensive and ratelimiting surgical domain experts. Herein, we demonstrate and validate a methodology to obtain high quality surgical tissue annotations through crowdsourcing of non-experts, and real-time deployment of multimodal surgical anatomy AI model in colorectal surgery

Introduction: Subjective surgeon interpretation of near infrared perfusion video is limited by low interobserver agreement and poor correlation to clinical outcomes. In contrast, quantification of indocyanine green fluorescence video (Q-ICG) correlates with histologic level of perfusion as well as clinical outcomes. Measuring dye volume over time, however, has limitations; it’s not on-demand, has poor spatial resolution, and is not easily repeatable. Laser speckle contrast imaging quantification (Q-LSCI) is a real-time, dye free alternative, but further validation is needed. We hypothesize that Q-LSCI will distinguish ischemic tissue and correlate over a range of perfusion levels equivalent to Q-ICG. Methods: Nine sections of intestine in three swine were devascularized. Pairs of indocyanine green fluorescence imaging and laser speckle contrast imaging video were quantified within perfused, watershed, and ischemic regions. Q-ICG used normalized peak inflow slope. Q-LSCI methods were laser speckle perfusion units (LSPU), the base unit of laser speckle imaging, relative perfusion units (RPU), a previously described methodology which utilizes an internal control, and zero-lag normalized cross-correlation (X-Corr), to investigate if the signal deviations convey accurate perfusion information. We determine the ability to distinguish ischemic regions, and correlation to Q-ICG over a perfusion gradient. Results: All modalities distinguished ischemic from perfused regions of interest; Q-ICG values of 0.028 and 0.155 (p<0.001); RPU values of 0.15 and 0.68 (p<0.001); X-corr values of 0.73 and 0.24 (p<0.001). Over a range of perfusion levels, RPU had the best correlation with Q-ICG (r=0.79, p<0.001) compared with LSPU (r=0.74, p<0.001) and X-Corr (r=0.46, p<0.001). Conclusion: These results demonstrate that Q-LSCI discriminates ischemic from perfused tissue and represents similar perfusion information over a broad range of perfusion levels comparable to clinically validated Q-ICG. This suggests that Q-LSCI might offer clinically predictive real-time dye-free quantification of tissue perfusion. Further work should include validation in histologic studies and human clinical trials.