Fluorescence angiography classification in colorectal surgery

Anastomotic leak is an important source of morbidity, mortality and rising healthcare costs Ashraf2013 in colorectal surgery. Surgeons use fluorescence angiography to optimise the perfusion of the tissue used to create the anastomosisKeller_ICG . The use of fluorescence angiography has been shown to reduce anastomotic leaks in non-randomized studies Blanco-Colino2018 ; Degett2016 and has been recommended by the European Association for Endoscopic Surgery [submitted]. The true utility of the technique remains unquantified.

The interpretation of the signal in fluorescence angiography is dependent on the surgeon and perhaps represents the actual ‘learning curve’ for this technique. These decisions have a high degree of variability shown both for static images [Soares, AS et al., Interobserver agreement study, under review] and dynamic video hardy2021inter assessments. Automating the interpretation of the signal would eliminate this variability and provide an alternative means of quantification, thereby leveraging the usefulness of surgical data science Maier-Hein2022 in the clinical setting.

Convolutional neural networks are especially suited for image analysis. The input for these neural networks are the pixel values of a given image and the outputs are the probabilities of the image belonging to a particular pre-defined class.

This paper describes the development of an artificial intelligence algorithm to automatically classify tissue based on perfusion using fluorescence angiography during colorectal surgery. A decision boundary is estimated based on the classification of different areas of the tissue of interest.

A dataset of patients undergoing colorectal resection was created with the characteristics shown in Table 1. Representative frames corresponding to the training set can be found in figure 1.

These results have show that a neural network trained on clinical data has achieved an accuracy over 90%. There was a drop in performance using data acquired with a different equipment for validation but still maintaining an accuracy over 80%.

The analysis of the saliency maps shows that in the cases where both false negative and positive results were obtained the pixels corresponding to the colon were not responsible for the main activation of the algorithm. This leads us to hypothesize that an image analysis pipeline segmenting the colon from the background first and, on this output, running the classification algorithm could lead to improved results. Based on previous work from our group [Soares, AS et al., Interobserver agreement study, under review] and others hardy2021inter significant variation exists between different surgeons’ assessments. Standardisation of fluorescence angiography is a necessary step to enable accurate comparison of results.

Our algorithm provides an easily accessible tool based on state of the art deep learning techniques that eliminates interobserver variability in interpreting the fluorescent signal. Previous efforts to standardise fluorescence angiography assessments have focused on quantitative measurements requiring post hoc signal processing wada2017icg ; son2019quantitative . Our algorithm has the ability to simplify post-hoc processing as it was trained on data captured intraoperatively and is designed to output a binary output (”fluorescent” vs ”non fluorescent”).

Some limitations should be addressed. Although a high number of frames was used for training, the variability of frames used for validation was high but the overall number was small. This should be mitigated by including data from more centres performing fluorescence angiography. Furthermore, the data collected was of low quality to begin with, which limited the usefulness of the algorithm. This should be interpreted with caution as even an experienced surgeon would find some trouble analysing this data intraoperatively. A close collaboration between engineers and clinicians has led to the development of the algorithm presented in this work. Further work should aim to enrich the dataset for further refinement of the algorithm. Clinical testing in a randomized clinical trial will be necessary before large scale implementation can occur.

An artificial intelligence algorithm to classify colonic perfusion based on fluorescence angiography was developed using state of the art deep learning techniques. This work paves the way for standardisation of fluorescence angiography in colorectal surgery.