Journal club: Ahmad O. et al, 2019

Artificial intelligence and computer-aided diagnosis in colonoscopy: current evidence and future directions


Lancet Gastroenterol Hepatol 2019; 4: 71–80


This paper reviews the different use of CAD (i.e. Computer Aided Diagnosis) and AI (i.e. Artificial Intelligence) in colonoscopy. In particular, it identifies three main areas of application in this field: polyp detection, polyp characterization, and procedure’s metrics identification.

As a matter of fact, colonoscopy is the gold standard for the screening of colorectal cancer, the fourth leading cause of death worldwide. However, the outcome of the procedure strongly depends on the ability and readiness of the operator. As a result of several studies, up to the 22% of adenomas (i.e. benign tumour with a potential of evolving into cancer) are missed in colonoscopies. In addition, it has been proved that having a second person, such as a nurse, watching the monitor during the procedure increases the ADR (i.e. Adenoma Detection Rate). For these reasons, CAD systems has been highly investigated as a tool for improving polyp detection. During the last decades many researches have focused on this topic, starting with machine learning approaches for the detection of polyp’s features (i.e. colour, edge, shape). Nowadays, the most promising results are given by CNN (i.e. Convolutional Neural Network) based approach. However, clinical trials for validating these systems are still needed. In addition, bigger datasets with high quality labelling, polyp morphology variability and multiple endoscopy processor are paramount for improving the deep learning algorithms.

A second area of application is the characterization of polyps as non-neoplastic (i.e. non cancerous) and neoplastic (i.e. potentially cancerous). Currently, almost all the potential tumoral masses encountered during colonoscopy are resected and sent to the histopathology unity for analysis. In situ characterization would allow to save costs and time, by allowing to resect only the dangerous masses. In addition, it would allow to design more appropriate treatments for the patients, based on the type of polyps found. Advanced imaging techniques have been investigated for this purpose. In particular, NBI (i.e. Narrow Band Imaging) magnification and endocytoscopy have given promising results. However, they rely on a technology which is not available in every endoscopy unity.

Finally, CAD and AI have been investigated for deriving online metrics about the quality of visualization, the cecal intubation rate, and to give a standardized score on the bowel preparation. As a matter of fact, current metrics used for assessing the quality of the colonoscopy are given only at the end of procedure. Doing so, they do not enable the operator to correct its actions if needed.


CAD and AI have the potential to overcome many difficulties related to colonoscopy. As a matter of fact, many studies have shown that they could assist the endoscopist and improve the diagnostic outcome of the procedure in terms of quality, time and costs. However, many challenges, listed below, need to be overcome:

  • Need of clinical trials, since almost all the researches in this field rely on retrospective studies. In addition, current results are obtained by testing the algorithms in controlled environments. Hence, understanding their performances in a real practise is needed.
  • Creation of large datasets including high quality labelling or annotation from experts based on pathology results.
  • Inclusion of additional information that clinicians use to provide the diagnosis (e.g. age of the patient, position of the mass, etc.).
  • Reduce the class imbalance of the datasets.
  • Overcome the limitation of using a non-transparent system for providing diagnostic outcomes.
  • Understand the position of the CAD in the colonoscopy workflow. Operator overloading with information, and potential distraction should be considered. Indeed, focusing on the information provided by the assisting tool, the doctor could potentially neglect other important data.
  • Understand the impact of false positive and false negatives in this field. As a matter of fact, although false positive (i.e. the system detects as a polyp a non-tumoral mass) could slow down the process, they have a minor impact than false negatives (i.e. the system does not detect a polyp).
  • Learn how to efficiently provide the CAD based knowledge intraoperatively avoiding to burner the endoscopist. Auditory feedback, for example, instead of augmented reality solutions, should be considered and appropriately tested.