PTU-058 Machine Learning Creates a Simple Endoscopic Classification System for Detecting Dysplasia in Barrett’s Oesophagus with i-Scan Imaging and Opens the Way to Standardised Training and Assessment of Competence

    loading  Checking for direct PDF access through Ovid

Abstract

Introduction

Barrett’s oesophagus (BE) is the pre-cursor for oesophageal adenocarcinoma. Endoscopic surveillance is performed to detect dysplasia in BE as it is likely to be amenable to curative treatment. Current surveillance relies on white-light endoscopy to obtain 4-quadrant biopsies through every 2cm of the BE segment. This samples less than 5% of the BE epithelium and is likely to miss dysplasia.

Introduction

A novel endoscopic image enhancement technology, i-Scan (PENTAX HOYA, Japan), has been developed to improve lesion recognition in the gastrointestinal tract (GIT). i-Scan uses post-processing light filtering to provide real-time analysis and enhancement of the mucosa and microvasculature.

Introduction

We evaluated the accuracy of i-Scan using a mucosal (M) and vascular (V) classification system for BE amongst 3 expert (consultant) endoscopists. Machine learning (ML) generates simple rules, known as a decision tree, to improve dysplasia detection and validate our classification system. To our knowledge, ML has never been applied for dysplasia detection in the GIT.

Methods

High definition video recordings were collected from patients with non-dysplastic (ND-BE) and dysplastic (D-BE) BE undergoing endoscopy at UCLH. A protocol was used to record areas of interest after which a matched biopsy was taken to confirm the histological diagnosis. In a blinded manner, videos were shown to 3 expert endoscopists who interpreted them based on their M and V patterns, presence of nodularity, ulceration and suspected diagnosis. Acetic acid (ACA) was used in some cases. Data was inputted into the WEKA package to construct a decision tree for dysplasia prediction.

Results

Videos from 47 patients (13 before and after ACA) were collected (24 ND-BE, 23 D-BE). Cases in which ACA was used, 7 had ND-BE and 6 D-BE. Experts’ average accuracy for dysplasia prediction was 72.2% (66.7–76.7%). ACA did not improve dysplasia detection. In 5 cases all 3 experts failed to detect D-BE.

Results

Using ML, the most important attribute was the lesions’ V pattern. If this was reported abnormal (irregular, dilated vessels) by more than one doctor, the lesion was D-BE (accuracy 79%). If D-BE was predicted despite the V pattern being reported abnormal by one or fewer experts, the lesion was still D-BE and vice versa.

Conclusion

Experts can diagnose D-BE in up to three-quarters of cases using i-Scan. ML can define rules learnt from expert opinion that predict dysplasia with a similar level of accuracy and are easier to learn than conventional classification systems. They could be used to train non-expert endoscopists in dysplasia detection and then used for blinded assessment of accuracy.

Disclosure of Interest

None Declared.

Related Topics

    loading  Loading Related Articles