Using AI to detect dysplasia in Barrett’s oesophagus

Barrett’s oesophagus develops when chronic gastro-oesophageal reflux changes the distal oesophageal lining. Dysplasia in this altered tissue can progress to oesophageal adenocarcinoma, one of the most serious consequences of long-standing reflux. Learn more about リウマチ患者の頸椎病変 環軸椎転位など の画像診断と定期フォロー.

Using artificial intelligence to detect dysplasia in Barrett’s oesophagus combines computer vision with endoscopic surveillance. The aim is to highlight subtle mucosal changes during upper gastrointestinal endoscopy, helping clinicians target areas that deserve closer inspection and biopsy.

ISDE 2021, held virtually from 27 to 30 September, brought together scientific sessions, workshops and research abstracts on oesophageal health. For Australian gastroenterologists following the congress, the practical challenge is translating imaging progress into a system where public and private pathways, pathology capacity and follow-up arrangements differ.

Why dysplasia can be missed

Dysplasia may appear as a barely perceptible colour or vascular change, particularly where the background Barrett’s lining is patchy. A competent endoscopist can still miss an isolated focus, especially when the examination is short or attention falls.

In a busy Australian public hospital, a waiting endoscopy list may stretch for months. Detection is therefore more than an image-processing problem: a suspicious area must lead to targeted biopsy, specialist review and a reliable plan for repeat surveillance.

What the AI actually sees

Deep learning systems are trained on annotated endoscopic images or video frames, often including white-light views and narrow-band imaging (NBI). Their software learns patterns associated with Barrett’s neoplasia, then produces a classification, segmentation or colour-coded risk map for the endoscopist.

Image quality remains decisive. Blur, bubbles, motion and inconsistent magnification can alter the cues a model relies upon, while false positives may prompt unnecessary biopsies. Its output is therefore a prompt for expert judgement, not a stand-alone diagnosis.

Most systems target visible neoplasia rather than proving invasion or grading dysplasia without tissue. This distinction keeps claims about image recognition separate from the pathology-based diagnosis that guides treatment.

Evidence that survives clinical scrutiny

A convincing evaluation needs external validation across hospitals, endoscopes and patient groups, with reference pathology and expert adjudication. Sensitivity, specificity, false-positive rates and time saved should be reported alongside comparisons with standard endoscopic practice.

A Japanese review of imaging diagnosis and follow-up for cervical lesions in rheumatology concerns a different specialty, but it reinforces a general principle: interpreting images matters most when the result is connected to review, confirmation and scheduled follow-up. Australian deployment likewise needs local evidence rather than imported performance.

Fitting AI into Australian endoscopy

Australia’s Medicare system covers eligible public hospital care and subsidises private gastroscopies through the Medicare Benefits Schedule, usually after referral from a general practitioner. There is no national population-screening program for Barrett’s, so most examinations follow reflux symptoms, other clinical indications or surveillance of known Barrett’s.

Hospitals from central Melbourne to regional northern Queensland operate different staffing and equipment arrangements. AI could provide a useful “second pair of eyes”, including for specialists working with remote teams, but telehealth cannot perform the biopsy and pathology still has to be completed locally. Clear referral and results-tracking pathways remain essential.

Privacy, regulation and responsibility

Clinical images and pathology data are sensitive health information. Australian privacy obligations—including the Privacy Act 1988 where applicable and relevant state health-records rules—govern collection, access, de-identification, storage and secure transfer, particularly when cloud services are used.

Software status also matters. The Therapeutic Goods Administration regulates medical devices, so a diagnostic AI tool requires the appropriate regulatory status for its intended Australian use rather than being treated as an ordinary reporting add-on. Procurement, cybersecurity, model updates, audit trails and clinician override need explicit ownership.

Turning detection into better outcomes

Detected dysplasia still requires histopathological confirmation. Visible suspicious lesions may be removed endoscopically, while confirmed flat high-grade dysplasia may be considered for ablation after specialist discussion. AI can improve targeting and reduce the number of random biopsies, but pathology workload and follow-up capacity determine whether that efficiency matters.

The strongest benefit may be fewer missed lesions rather than a visibly faster procedure. Studies should therefore track interval cancers, repeat examinations and treatment as well as image-level accuracy. Australian patients also need clear explanations of what the software contributes, what remains uncertain and who is responsible for the final decision.

Requirements for routine clinical use

Australian implementation should begin with defined endpoints, local audit standards and a plan for reviewing disagreements between the model, endoscopist and pathologist. Cost-effectiveness differs between a metropolitan day-surgery unit and a regional service where extra travel and staffing may outweigh modest efficiency gains.

ISDE’s emphasis on scientific exchange and continuing education provides a useful foundation for shared datasets, reproducible standards and post-deployment surveillance. Routine use still depends on prospective evidence that patient outcomes improve without widening disparities in access.

The practical takeaway is a simple rule of thumb: AI is useful in Barrett’s when it reliably directs attention, preserves tissue for confirmation and fits an Australian follow-up pathway. No heat map can replace pathology, multidisciplinary judgement or timely treatment.

For hospitals and specialists, the immediate gain is a disciplined workflow built around those requirements, with measured performance and clear accountability. That keeps each flagged finding tied to confirmation and care while leaving the final decision firmly with the treating team.