Computer Aided Diagnosis in Upper GI Endoscopy (AIOGD)

This study aimed to investigate the clinical application of Artificial Intelligence Software for computer aided diagnosis (CAD), for real-time anatomical coverage, automatic Identification, classification and interpretation of abnormal lesions in upper GI endoscopy, and benchmarking their accuracy compared to endoscopists.

5.1 Phase One This phase of the study is non-interventional. It involves two sections: training cohort and validation cohort. 5.1.1 Training Cohort Such artificial intelligence models typically need a lot of data to be fed in, so that it can predict the results with a good accuracy. Since it uses machine learning, the accuracy gets better with each data set that is fed in. For the training cohort, prospective collection of videos on diagnostic OGD would be done. Selected endoscopists would be primed on the examination protocol. Patients would be sedated and given oral N-acetylcysteine 30 mins before the procedure. Investigators will record basic epidemiological, analytical and clinical data from patient records. ICD-10 diagnostic codes or similar diagnostic recording measures and recorded lab data will be used to determine diagnosis. The Procedure will be done using Olympus 290 systems under conscious sedation after informed written consent. The procedure will be recorded and the video file will be saved for future use. Images of suspected lesions and certain anatomic landmarks will also be saved like Esophagus (upper, middle, lower), gastroesophageal junction, fundus, body and antrum of stomach, incisura and pylorus and first and second part of duodenum. These images will also be used in initial training of the CADT. Since all endoscopic procedures in the endoscopy centre of Prince of Wales Hospital have been recorded since December 2018, some OGD that has been performed will be screened for quality and would also be used in this training cohort. 150-200 patient videos would be marked at different anatomical sites by trained endoscopists. These videos would be anonymized. The training cohort would be repeated until the sensitivity and specificity are over 90%. 5.1.2 Validation Cohort After feeding these images to the software, a validation cohort would be performed. All consecutive patients undergoing diagnostic OGD meeting the inclusion and exclusion criteria in the Prince of Wales Hospital would be included in the study. The endoscopists would be blinded to the presence of the CADT and the standard protocol. The endoscopist would perform a diagnostic OGD as per usual practice. The procedure would be recorded and accessed by the CADT at the same time. The same video would be validated by a specialist endoscopist on the adequacy of the anatomical sites covered and the preparation of the stomach. This would be considered as the gold standard. The specialist endoscopist would be blinded to the result of the CADT. The CADT result would be compared against the specialist endoscopist. Sensitivity and specificity of the CADT results would be generated. If it is lower than 90%, further pictures/ videos have to be fed in for the training cohort.

Computer Aided diagnosis is an environment-enhancing technology developed to aid humans to improve their accuracy, speed and confidence by providing live feedback during the procedure, on request, or after the procedure has been completed.

Inclusion Criteria: Patients undergoing diagnostic OGD for evaluation of their symptoms Age >= 18 Exclusion Criteria: Patients refusing consent/not willing to participate in the study Patients where biopsies cannot or have not been taken Patients with history of gastrectomy Patient admitted for upper gastrointestinal bleeding Patient who are not suitable for EGD examination (such as acute peritonitis with suspected perforated bowel) Pregnant females