AUGUR-AI - Indocyanine Green Fluorescence Angiography Representer

Surgery is a major contributor to healthcare, but it must be continually optimised to maximise its effectiveness while minimising risk and cost. Complex surgical outcomes, which often require significant healthcare resources, highlight the need for constant improvements in safety and precision. Every surgical procedure consists of a series of critical steps that rely on surgeon judgement as well as technical dexterity and skill. Decisions are made predominantly through visual interpretation of findings in the context of patient specific factors (including those profiled preoperatively) and the operator's own experience (which includes formal training and accumulated expertise). This is especially true for minimally invasive surgery which uses a camera placed internally to display the internal scene along with the operator's actions on a screen for viewing (the digitised image is created by a computer for interpretation by a human). Crucial to the success of the operation is making the most accurate decisions as possible to ensure the patients gets a safe, successful operation that cures the problem with as little functional loss as possible. For colorectal resectional surgery, this means knowing the residual tissues are sufficiently perfused to heal, as non-healing of anastomoses is associated with significant morbidity and mortality.

Near-infrared laparoscopy is an established technology that aims to improve clinical outcomes by providing a means of seeing certain tissue characteristics that would otherwise be invisible to the human eye to aid intraoperative decision-making. This technology uses energy in the NIR range (780-820nm) to visualise tissues alongside standard white light interrogation. Energy in this wavelength causes no cellular damage and can penetrate some tissue to a depth of some millimetres. At this wavelength, there is no biological reflectance from the tissue so presence of a responsive agent administered systemically can be confirmed if the agent is capable of fluorescence (that is absorption of energy at one wavelength and its re-emission at a different wavelength). While new fluorescence agents are in development, right now the most useful and safest agent approved for use is indocyanine green (ICG), which has long been approved and proven safe for circulatory assessment including microvascular tissue perfusion. When administered systemically, the dye circulates through the body in seconds, providing real-time information that can be repeatedly assessed at various stages throughout the surgery. To date, many studies, including randomised controlled trials, have demonstrated that the use of indocyanine green fluorescence angiography (ICGFA) is associated with lower rates of anastomotic leak in major colorectal resectional surgery. Subjective interpretations of dynamically changing scenes can be difficult, especially where differing areas of the screen may need to be tracked over time. However, given that the use of ICGFA is associated with lower leak rates, interpretation of signal patterns by expert users must be consistent and this has also been demonstrated in work looking at inter-user variability.

"Digital surgery" as a concept relates to the application of technology for real-time data analytics during operations. The term Artificial Intelligence (AI) refers to technology that attempts to mimic human cognition. Computer vision refers to the conversion of visual display into numeric datasets than then become available for machine learning via its application of statistical models based on a sample of data to apply decision-making to situations without being explicitly designed to perform the task. A predictive model can weigh different paths and a classifier can then present decision support suggestion alongside its confidence levels. In previous work, we have trained AI models based on expert interpretation of fluorescence signal patterns in patients who were known not to suffer a post-operative anastomotic leak. We have subsequently tested this method prospectively, intra-operatively in consecutive patients on a laptop (with NVIDIA T500 2GB GPU). The software method proved 100% accurate at object-level and matched the actual stapler site placement by the operating surgeon. We have also shown that it can learn from interpretations of multiple surgeons and is generalizable to other surgeons and imaging systems. Since then, we have built a software platform on NVIDIA Jetson.

We aim to test the AI-based software in providing how an expert surgeon would interpret ICGFA signals in real time, in-theatre during surgery. Surgeons will be blinded to the predictions made by the AI algorithm.

To do this, ICGFA will be used as approved in patients undergoing open, laparoscopic or robotic major colorectal resectional surgery as per usual clinical care protocols. Video recordings will be analysed by the AI model alongside the operation, however the surgeon will be blinded to the representation.