P143 Performance of a convolutional neural network for Barrett's oesophagus related dysplasia detection following quality labelling. (19th June 2022)

• Record Type: Journal Article
• Authors: Aslam, Nasar; Hussein, Mohamed; Kader, Rawen; Mejias, Anton; Brandao, Patrick; Puyal, Juana Gonzalez-Bueno; Mountney, Peter; Stoyanov, Danail; Lovat, Laurence; Haidry, Rehan
• Is Part Of: Gut. Volume 71(2022)Supplement 1
• Page Start: A110
• Page End: A111
• Publication Date: 2022-06-19
• Subjects: Gastroenterology -- Periodicals; 616.33
• DOI: 10.1136/gutjnl-2022-BSG.198
• Languages: English
• ISSNs: 0017-5749

Abstract

Abstract : Introduction: Invasive oesophageal adenocarcinoma arising in Barrett's oesophagus (BE) can occur in up to 25% of patients one year after index endoscopic examination. Computer assisted diagnosis may reduce the incidence of overlooked early neoplasia in BE patients. Our group has previously trained a convolutional neural network (CNN) that achieves dysplasia detection rates comparable to expert endoscopists. The aim of this study was to evaluate the effect that image quality has on the performance of our CNN. Methods: We previously trained a CNN with Resnet101 architecture to classify videos frames according to the presence or absence of dysplasia. All available video frames demonstrating dysplasia from high-definition white light and i-scan Pentax (Hoya, Japan) imaging modes were used as part of the training for the CNN. Histology from mucosal resection specimens or biopsies served as the ground truth for the presence of dysplasia. Low- or high-quality markers were assigned to each frame with dysplasia present. Quality labelling was assigned by a single reviewer. Factors that influenced quality labelling included overall visibility of the dysplastic lesion, motion blur, light speculation, or artefact. Results: A total of 23513 frames of dysplastic BE were assigned quality labels across 17 cases. 17915 frames were labelled as high quality with 5598 being of low quality. Evaluating the CNN performance solely on high quality frames yielded an overall sensitivity of 84.61%. When testing against both high- and low-quality frames the CNN achieved an overall sensitivity of 82.26% ( table 1 ). Conclusions: Evaluating the performance of our CNN on a combination of high- and low-quality dysplastic frames yields comparable overall sensitivity to testing solely on high quality dysplastic frames. Real time endoscopy captures a mix of high- and low-quality frames which is influenced by factors such as artefact, patient tolerance and scope stability. The comparable performance of our CNN will be valuable for future real time application.