AS Soares, MSc, S Bano, D Stoyanov, Professor, LB Lovat, Professor, M Chand, Ass Professor. Wellcome / EPSRC Centre Interventional and Surgical Sciences
Objective of the technology
This computer vision algorithm provides a standardisation of fluorescence-guided perfusion assessment in anterior resection of the rectum.
Description of the technology
Background – Assessment of perfusion of the left colon with fluorescence during anterior resections for cancer changes surgical decisions in up to 19% of cases. The use of fluorescence in this setting has been shown to be associated with lower leak rates, and improved short- and long-term outcomes with reduced costs. Given the high incidence of colorectal cancer, fluorescence-guided perfusion assessment could be of great importance in contemporary surgical oncology practice. However, there is currently no standardisation of this technique which represents a significant limitation to widespread adoption.
Data – video data were collected by a single surgeon in a referral centre for colorectal cancer treatment. Perfusion assessment was used before proximal colon division to identify the best location for transection. A bolus of indocyanine green was injected intravenously and a near infrared camera was used to assess perfusion of the area of interest through fluorescence.
Algorithm Development – Photographs of fluorescent imaging of the colon were analysed using a non-supervised learning algorithm called “K-means clustering”. The first step was to digitally subtract all background pixels, leaving only the area of interest of the colon. The second step was to subsegment this into 2 “clusters” corresponding to perfused and non-perfused areas. A mathematical model was then proposed based on the 2 sub-clusters centres to select the area for transection with optimal perfusion of the proximal colon.
Preliminary results
Representative images of proximal colon under perfusion assessment with fluorescence were presented to 2 expert surgeons. The optimal point for transection was then selected based on their clinical judgement on previously delimited areas indicated by random letters (see fig. 1). This was compared with the results from the automated segmentation using the computer vision algorithm (see fig. 2). The area identified for section by the algorithm included the area selected by the expert surgeons in 100% of the test cases.
fig. 1 – possible areas for transection presented to expert surgeons
fig. 2 – decision boundary overlay on perfused colon
Conclusions and future directions
These results need to be further externally validated, as the risk of overfitting in this experiment is high. The next steps include the collection of multicentre data with a standardised fluorescence perfusion assessment. Further areas to explore include the development of a deep learning algorithm for perfusion segmentation using the multicentre data.
After robust training, the algorithm will be validated on real-time clinical data to ensure improved outcomes for patients, which is the ultimate goal of this technology.
Presented at the SAGES 2017 Annual Meeting in Houston, TX.
Abstract ID: 98895
Program Number: ET003
Presentation Session: Emerging Technology Session
Presentation Type: Podium