P1.140. A Computer Vision Model for Anatomic Recognition in Robotic Assisted Minimally Invasive Esophagectomy

Abstract

Esophageal Cancer: Other Esophagectomy is associated with relatively high complication rates. Computer vision models assisting trainee surgeons in the recognition of critical anatomy have the potential to reduce complications. Our lab is working to design a computer vision model capable of recognition of anatomy pertinent for robotic assisted minimally invasive esophagectomy (RAMIE). RAMIE videos were recorded on the DaVinci Surgical System. In order to increase heterogeneity of the dataset while maintaining a manageable annotation burden, clips of critical portions of the procedure were selected from each video and extracted at 1 frame per second and uploaded to the computer vision annotation tool. Critical structures including lung, diaphragm, esophagus, azygos vein, stomach/gastric conduit, trachea, and lymph nodes were chosen for annotation. A team of undergraduate and graduate students assisted in the annotations and underwent training in relevant anatomy and annotation technique by a surgical resident prior to being allowed to annotate. Annotations were reviewed on a weekly basis for accuracy and annotators were retrained if needed. A U-Net model was designed in python and trained on our growing surgical video dataset. Dice score was used to evaluate the performance of the model across each label classification. The designed model was able to achieve a Dice score averaging 70.09% across three label classes on a publicly available pre-annotated dataset. When testing on the growing surgical dataset (currently ~2500 frames), there is a high variability in the Dice scores for each structure ranging from 0.1-0.80 with steady improvement throughout training. The Dice score was proportional to the representation of each class in the data with the model experiencing more difficulty recognizing less frequently appearing labels such as the trachea. The Dice scores improved both across training epochs and with the addition of more annotated frames to the surgical video dataset. The results obtained on public datasets suggest viability of the model, and the high variability of the results obtained from testing on our own data is explainable by the currently small sample size and complexity of the model This is expected to improve as more data becomes available. With further improvement the model may serve as a useful adjunct to surgical trainees to help with early identification of critical structures to avoid inadvertent injury.