P1.141. A Machine Learning Model for Assessment of Conduit Perfusion in Robotic Assisted Minimally Invasive Esophagectomy Using Indocyanine Green

Abstract

Abstract Topic Esophageal Cancer: Other Background Anastomotic leak remains a major complication in esophagectomy. Indocyanine Green (ICG) can assist with the assessment of conduit perfusion but remains subjective and prone to bias. Our lab has worked to design a computer vision model capable of judging the presence or absence of ICG within a gastric conduit. Methods Videos of robotic assisted minimally invasive esophagectomies (RAMIE) were recorded in their entirety using the DaVinci Surgical System. Segments where ICG was used were uploaded as clips to the computer vision annotation tool (CVAT). A team of undergraduate and graduate students underwent training in annotation technique by a surgical resident, and portions of the conduit where ICG could be visualized were hand annotated resulting in a total of 248 frames. All annotations were reviewed for accuracy prior to inclusion in the dataset. Using a U-Net architecture, a computer vision model intended for binary image segmentation was designed in house and tested first on a large publicly available dataset (approximately 3000 frames) to ensure learning capability. The model was then trained instead on ICG annotations and common machine learning metrics including Intersection over Union (IoU), and Dice score. Results The model was able to achieve a Dice score of 0.90 on a publicly available dataset intended for binary image segmentation. When tested on our growing surgical video dataset, the model was able to achieve a Dice score of 0.43 with similar training parameters with a high variability in the IoU between individual images. These results were converged upon after approximately 60 training cycles after which no further improvements occurred. The model was able to generate predicted masks for visualization purposes consistent with calculated IoU. Conclusion Given the model demonstrated improvement across training epochs on both datasets, the poor performance of the model on the ICG dataset compared to the larger dataset is likely a function of the small sample size and is expected to improve as our lab continues to collect further video data. Future steps include the addition of intensity and time to fluorescence as well as correlation with patient outcomes to create an objective tool for anastomotic assessment.