QS102 - AI-Technology Innovations in Robotic Colorectal Surgery – Detecting Optimal Angles for Critical Exposure and Division of the Inferior Mesenteric Artery

Image or Table (Do Not add text here):

Purpose/Background: As with many aspects of surgery, operational outcomes are highly dependent on adherence to evidence-based guidelines and techniques. With the recent growth of robotic-assisted surgery, there is a unique opportunity to analyze videos of every surgical maneuver made. These videos have a wealth of untapped data which is paramount to optimizing surgical techniques.

A critical step in Low Anterior Resection (LAR) and sigmoidectomy is the division of the Inferior Mesenteric Artery (IMA). This step requires difficult surgical maneuvers including dissection, exposure, and optimal instrument positioning. Despite the importance, the optimal retraction and relative tissue relationship metrics for critical exposure during IMA division have yet to be defined. Our work leverages high-resolution intraoperative videos and artificial intelligence (AI) to capture and quantify key angles and approaches used by experienced colorectal surgeons.

Methods/Interventions: Deidentified videos of robotic-assisted LAR and sigmoidectomy surgeries performed by colorectal surgeons were analyzed. Videos were manually annotated for segmentation using a 5-15 second time interval in the moments during and preceding stapler clamping. Video segments were analyzed with Principal Component Analysis (PCA) and automated geometric object identification. PCA identified two principal axes for each segment and defined two directional vectors corresponding to the IMA and stapler (Figure 1). Using these vectors, three angles were extracted: 𝑎, the angle between IMA vector and horizontal lower boundary of frame; 𝑏, the angle between stapler vector and vertical boundary of frame; and 𝑐, intersection angle between IMA and stapler vectors.

Results/Outcomes: AI-analysis was completed for nine robotic surgery videos performed by two colorectal surgeons at two hospitals. During IMA division, the mean and standard deviation across all videos for angles 𝑎, 𝑏, and 𝑐 (Figure 1) were 58.4° (14.7°), 130.4° (15.8°), and 86.2° (10.9°), respectively. These values were consistent among different surgeons with a two sample t-test unable to detect a difference for any angle (𝑎, 𝑏, or 𝑐) between surgeons (p-values 0.12, 0.27, and 0.31, respectively). Among all frames for each video (25-75 frames per video captured at 5 frames per second), the AI software identified the mean angles across all frames with extreme precision (standard deviation ˂1° for each angle).

Conclusion/Discussion: We created an AI-model to capture optimal angles and relative instrument relationships during IMA division among a small sample of colorectal surgeons. Results showed consistency of intraoperative angles across different videos and between surgeons. With minimal manual labor, our model can process 5-15 second clips into hundreds of frames to generate unbiased, intraoperative data. We are in the process of collaborating with surgeons at other institutions to obtain upwards of 500 videos to be analyzed with our model.

With no established intraoperative guidelines for robotic surgery, we aim to capture the vast operative data from experienced colorectal surgeons. By systematically analyzing their surgical maneuvers, we transform their instinctual operative knowledge into quantifiable measurements. These metrics can ultimately be used for optimizing intraoperative maneuvers, educating new surgeons on correct technique, and retrospectively analyzing complications. As this is the first of many projects, we are also working on a model for tension estimation in colorectal anastomoses.