{
  "abstract": "Introduction Barrett’s oesophagus related neoplasia (BORN) can be inconspicuous or subtle on white light endoscopy (WLE), therefore random biopsies are recommended. Artificial intelligence can aid in the detection of BORN. The aim of this study is to evaluate the diagnostic accuracy of a commercial AI system in comparison to expert endoscopist assessment.Methods This was a single centre prospective study. Eligible patients had at least 1cm of Barrett’s oesophagus (BO). The Barrett’s was inspected first by the endoscopist using WLE and narrow band imaging (Olympus Evis X1). Near focus was used for characterization lesions, which were classified as high or low level of suspicion. Odin Vision CADU system was then turned on, which highlights suspicious lesion within a green bounding area. A positive AI signal was classified in ‘consistently positive’ or ‘transient positive’ depending on the stability of the bounding area. Any lesion received at least one targeted biopsy, before Seattle protocol biopsies. Sensitivity, specificity, negative predictive value (NPV) and positive predictive value (PPV) for high-grade dysplasia (HGD)/ intramucosal adenocarcinoma (IMC) were assessed at different thresholds of detection. McNemar’s test was conducted to compare CADU to endoscopist assessment. Per-lesion and per-patient analyses were performed separatelyResults A total of 68 patients were enrolled in the study, for a total of 140 areas of interest (IMC=18; HGD=16; LGD=8; Barrett’s without dysplasia n=98).Two thresholds for detection were used: the stringent threshold required a ‘consistently positive’ AI signal or high level of endoscopist suspicion, while the lower threshold was met by a ‘transient positive’ AI signal or low level of endoscopist suspicion. The sensitivity, specificity and diagnostic accuracy levels are shown in table 1.There was no difference in sensitivity of the CADU and endoscopist assessment at both thresholds in both the per-lesion and per-patient analysis. There was a statistically significant difference in specificity in the per-lesion analysis at the lower threshold (p<0.0001), but not at the higher threshold or in the per-patient analysis.Conclusion CADU system has good sensitivity for HGD and IMC, with performance comparable with expert endoscopists. Lower threshold of detection maximizes sensitivity, but suffers from low specificity. Future work should compare CADU with non-expert endoscopists.Abstract O29 Table 1Per-lesion analysisPer-patient analysisStringent ThresholdLower ThresholdStringent ThresholdLower ThresholdCADUEndoscopistsCADUEndoscopistsCADUEndoscopistsCADUEndoscopistsSensitivity0.730.820.880.970.780.810.890.89Specificity0.570.690.380.170.430.540.200.07PPV0.360.460.310.270.450.510.380.36NPV0.870.920.910.950.770.830.750.50",
  "authors": [
    {
      "affiliations": [
        "Cambridge University Hospitals NHS Trust, Cambridge, UK"
      ],
      "name": "Audrey Fu"
    },
    {
      "affiliations": [
        "Cambridge University Hospitals NHS Trust, Cambridge, UK",
        "Early Cancer Institute, University of Cambridge, Cambridge, UK"
      ],
      "name": "Andreas Hadjinicolaou"
    },
    {
      "affiliations": [
        "Cambridge University Hospitals NHS Trust, Cambridge, UK"
      ],
      "name": "Vijay Sujendran"
    },
    {
      "affiliations": [
        "Cambridge University Hospitals NHS Trust, Cambridge, UK",
        "Early Cancer Institute, University of Cambridge, Cambridge, UK"
      ],
      "name": "Massimiliano di Pietro"
    }
  ],
  "title": "O29 Artificial intelligence versus expert endoscopists for the detection of barrett’s-related neoplasia: a single centre prospective study",
  "uid": "94c8f7d1-95dd-52f1-892a-9d6fc2c760cc"
}