Advertisement

Development and Validation of Artificial Neural Networks Model for Detection of Barrett’s Neoplasia, a Multicenter Pragmatic Non-Randomized Trial

Open AccessPublished:October 22, 2022DOI:https://doi.org/10.1016/j.gie.2022.10.031
      This paper is only available as a PDF. To read, Please Download here.

      ABSTRACT

      Background & aims

      The aim of this study is to develop and externally validate a computer aided detection (CADe) system for detection and localization of Barrett’s neoplasia and assess its performance compared to general endoscopists in a statistically powered multi-center study using real-time video sequences.

      Methods

      In phase-1, the hybrid VGG16-SegNet model was trained using 75,198 images and videos (96 patients) of neoplastic and 1,014,973 images/videos (65 patients) of non-neoplastic Barrett’s. In phase-2, image-based validation was performed on a separate dataset of 107 images (20 patients) of neoplastic and 364 images (14 patients) of non-neoplastic Barrett’s. In phase-3 (video-based external validation) we designed a real-time video-based study with 32 neoplastic videos (32 patients) and 43 non-neoplastic (43 patients) Barrett’s videos from four European centers to compare the performance of the CADe model to that of six non-expert endoscopists. The primary end point was the sensitivity of CADe diagnosis of Barrett’s neoplasia.

      Results

      In phase 2, CADe detected Barrett’s neoplasia with sensitivity, specificity and accuracy of 95.3%, 94.5% and 94.7% respectively. In phase 3, the CADe system detected Barrett’s neoplasia with sensitivity, specificity, NPV and accuracy of 93.8%, 90.7%, 95.1% and 92.0% respectively compared to the endoscopists’ performance of 63.5%, 77.9%, 74.2% and 71.8% respectively (p<0.05 in all parameters). The CADe system localized neoplastic lesions with accuracy, mean precision and mean IoU of 100%, 0.62, 0.54 respectively when compared to at least one of the expert markings. The processing speed of the CADe detection and localization were 5ms/image and 33ms/image respectively.

      Conclusion

      This is the first study describing external (multi-center) validation of AI algorithms for Barrett’s neoplasia detection on real-time endoscopic videos. The CADe system in this study significantly outperformed non-expert endoscopists on real-time video-based assessment achieving >90% sensitivity for neoplasia detection. This needs to be validated during real-time endoscopic assessment.
      Acronyms and abbreviations: LGD (Low grade dysplasia), HGD (High grade dysplasia), EAC (Early adenocarcinoma), ASGE (American society of gastrointestinal endoscopy), AI (Artificial intelligence), REC (Research ethics committee), CNN (Convolutional neural network), CADe (Computer aided detection), VGG (Visual geometry group), FPS (Frame per second), WLI (White light imaging), GOJ (Gastro-oesophageal junction), CLAHE (Contrast Limited Adaptive Histogram Equalization), IoU (Intersection over union), SD (Standard deviation), NPV (Negative predictive value), AUC (Area under the curve), ROC (Receiver operator curve), PIVI (Preservation and Incorporation of Valuable endoscopic Innovations), CADx (Computer aided diagnosis (characterization))