AI in colonoscopy and beyond: On the cusp of clinical implementation?

The ongoing excitement around artificial intelligence (AI) applied to healthcare stems from the ‘deep learning’ revolution. Deep learning refers to a subset of AI techniques that use neural networks with complicated structures to perform tasks, often related to analysis of images of text.

We have come a long way since the first study applying deep learning to medicine, which looked at identifying whether skin lesions are cancerous. ¹ Two years later, we witnessed the first randomised control trial (RCT)—this time in gastroenterology. Wang et al. found that AI‐aided detection of polyps in colonoscopy led to greater detection of diminutive adenomas. ² Since then, gastroenterology has continued to lead the field, with five further RCTs—more than all other specialties combined.

In this issue, Hann et al. outline what these trials have shown and consider the proximity of clinical application of such algorithms. ³

For endoscopic lesions, AI has been explored to detect lesions (CADe) and characterise them (CADx). As Hann et al. highlight, all prospective trials have looked at the former, and AI is consistently found to detect more small polyps. ³ Some studies suggest that AI‐aided techniques could yield a 50% increase in adenoma detection rate and can reduce colonoscopy‐related costs up to 20%. ⁴ This appears promising, but further studies are still required to investigate whether this increased detection translates to reduced rates of cancer and increased survival.

The need to measure defined clinical endpoints (such as survival rates), and not just intermediates (such as polyp detection rate), is an important requirement before the widespread implementation of any AI algorithm in health care. Such endpoints should include measures of harm. With colonoscopy, false positives could lead to unnecessary polypectomies and longer procedures.

These endpoints must be measured in prospective trials. Research has highlighted promise of AI across gastroenterology, from analysing liver ultrasound images ⁵ to prognosticating in hepatocellular carcinoma, ⁶ and from personalising pancreatic cancer management ⁷ to predicting liver transplantation survival. ⁸ However, the impact of incorporating such algorithms into clinical workflows has not yet been robustly assessed. ⁹ Algorithms that perform well in retrospective studies can have a negative effect when implemented, as pointed out by the European Society of Gastrointestinal Endoscopy (ESGE). ¹⁰ We saw an example of this in CADe for breast cancer. ¹¹ AI for colonoscopy is leading the field in this respect, but there is still further work to be done.

Another important step is replicability across different settings. In CADe for colonoscopy, five out of the six RCTs to date were performed in China, and all were single‐centre studies. To be confident that findings translate elsewhere, we need large, multi‐centre prospective trials conducted in different countries. It is feasible that variations in mucosal appearance, polyp morphology and endoscopy technology could affect this generalisability. ESGE has reinforced this sentiment by recommending CADe and CADx only when ‘reproducible accuracy for colorectal neoplasia is demonstrated in high‐quality multi‐centre, in vivo clinical studies’. ¹⁰ Other application areas will need the same.

For many groups, the development of AI algorithms for health care represents a commercial opportunity. Hann et al. outline six such commercial CADe systems for colonoscopy. ³ As well as developing robust evidence of effectiveness, such algorithms must meet regulatory approval. The United States Food and Drug Administration is currently treating AI algorithms as medical devices. However, these algorithms present additional challenges. AI models can be updated after deployment in ways that traditional medical devices cannot. Heterogeneity in data collection methods and clinical workflows can also impact performance. The regulatory landscape will thus need to adapt. Following regulatory approval, commercialisation would require an appropriate method of reimbursement to be devised. To our knowledge, the only current precedent is Vizai's stroke detection algorithm, which was recently granted reimbursement from Medicare in the United States.

The strengths of AI align well with the demands of colonoscopy. The speed of AI computation allows real‐time input and deep learning specialises in image analysis. Colonoscopy is doctor‐led, naturally embedding strong oversight of the algorithm. These factors, amongst others, have enabled colonoscopy to lead the field regarding evidence of positive clinical impact. This establishes an important role in laying the path towards implementation that other AI algorithms may follow. Future advances will be driven by those with both clinical expertise and technological understanding. ¹² There is still work to be done, in particular multi‐centre, international RCTs, tailored regulatory frameworks and a method of reimbursement, but the widespread clinical implementation of our first AI algorithms is inching ever closer.

CONFLICT OF INTEREST

The authors have no conflicts of interest to declare.