Deep Learning Assisted Phase Recognition in Endoscopic Third Ventriculostomy Procedures

Endoscopic Third Ventriculostomy (ETV) is a low-complication surgical procedure for treating obstructive hydrocephalus. The complications that occasionally occur are avoided by following recommended practices and accurately handling surgical tools. Automated detection of instruments employed and landmark segmentation is expected to aim the surgeons to follow the protocol precisely, leading to a successful outcome.Deep learning is one of the best-automated techniques to extract tool-related information and surgical-phase recognition through landmark detection technique. In this work, two models are applied to conduct phase recognition for ETV surgeries that includes UNet (with a Resnet-152 backbone) with additional post-processing steps for anatomical landmark segmentation and YOLOv8 for instrument detection. These models are combined to form an ensemble that predicts the ETV phase for a given frame. The five phases of an ETV surgery include i) In the Lateral Ventricle, ii) Entering Third Ventricle, iii) In the Third Ventricle, and iv) Puncturing, and v) Dilation. The proposed ensemble model (UNet+YOLOv8) performs efficiently on the unseen test videos demonstrating an accuracy of 96% on average. Additionally, the models are reported to generate faster outcome and are more robust than the other baseline models that are investigated. The autonomous phase recognition in an endoscopic neurosurgery is the first of its kind and is a step towards making robotic assisted ETV surgeries a reality.