There is a growing need for cost-saving strategies that enable optical diagnosis of colorectal polyps to reduce the patient, economic, and environmental burden of polypectomy and pathology. One recommendation has been to adopt a leave-in-situ approach for small, nonthreatening hyperplastic rectosigmoid polyps ≤5 mm and a resection-and-rule approach for more proximal lesions ≤5 mm.
Overall, these tiny lesions account for more than 80% of all polyps detected during screening and surveillance colonoscopy, and an on-site diagnosis and treatment strategy and a resection and discard strategy could generate millions of dollars in savings related to endoscopy.
Against this background, Cesare Hassan, MD, PhD, of Humanitas Research Hospital in Rozzano, Italy, and colleagues tested the predictive diagnostic value of artificial intelligence (AI) versus standard histology for diagnosing colorectal polyps in a study from the non-random real life, which was recently published on Clinical Gastroenterology and Hepatology. Hassan discussed the results of the study in the following interview with the Reading room.
What was the impetus for your group to undertake the study of AI?
Hassan: The main objective was to test in a real-life environment the performance of an AI machine that could predict the histology of polyps in standard white-light endoscopy.
What had previous research shown about the potential for an AI approach to colorectal polyps?
Hassan: The application of AI to colonoscopy actually began with the pioneering study by Dr. Yuichi Mori, which showed the possibility of predicting the histology of polyps using a magnified endoscope with the use of advanced imaging. However, there was uncertainty as to whether similar results could be achieved with a simple standard endoscope without the use of advanced imaging.
Unexpectedly, the implementation of AI has been hampered in part by suboptimal accuracy reported by the endoscopy community.
What did your group’s findings add to the picture?
Hassan: Our study showed in a real-life scenario the feasibility and accuracy of AI in predicting polyp histology, equaling the limit required for its clinical implementation.
Computer-aided diagnosis (CADx) without advanced imaging exceeded the benchmarks required for optical diagnosis of colorectal polyps. This approach could help implement cost-saving strategies in colonoscopy by reducing the burden of polypectomy and/or pathology.
Overall, 544 polyps were removed from 162 patients, of which 295 were histologically verified rectosigmoid lesions ≤5 mm. The diagnosis of CADx was feasible in 291 and the negative predictive value for rectosigmoid lesions ≤5 mm was 97.6%.
Of these 295 lesions, 242 were amenable to a leave-in-situ strategy, while 212 of the total 544 would have been amenable to a resection and discard strategy. This resulted in a 95.6% concordance between CADx-based and histology-based surveillance intervals according to European and American guidelines, respectively.
What are the immediate implications of the findings?
Hassan: Unlike AI for polyp detection, implementation of AI-assisted optical biopsy will be complex. AI assistance is likely to be increasingly used for the leave-in-situ strategy for non-neoplastic polyps in the rectosigmoid, standardizing a practice that is already widespread among endoscopists. On the other hand, the implementation of strategies based on the discarding of the neoplastic polyp after histology will face several non-AI barriers that still force most endoscopists, and more so general practitioners, to send excised lesions to histology .
Do you think the gastroenterology community will quickly switch to a profitable approach to AI?
Hassan: No doubt about that! Endoscopy is performed in real time and the risk of human error is very high. No endoscopist would reasonably refuse the help of a dedicated machine for the good of their practice and their patients.
What are the limitations of your study?
Hassan: Mainly, the fact that it was not randomized. Therefore, we can assume that the endoscopist was overconfident because he knew that the polyps would eventually be resected.
What questions still remain to be answered?
Hassan: We still regard histopathology as the diagnostic gold standard, and a substantial risk of misclassification that would unfairly penalize AI has been reported. Presumably, the use of AI also for pathology can further reduce the uncertainty related to research in this field.
What is the overall message of this research?
Hassan: AI can compensate for the limitations of endoscopic technique and technology. Using the same standard white-light endoscopy that most endoscopists use in their daily routine, an AI machine can predict polyp histology to the same extent as the most skilled endoscopists in our field, which marginalizes further the additional value of pathology in our practice.
You can read the abstract of the study hereand on the clinical implications of the study here.
This study received no funding.
The authors had no conflicting interests to declare.