S558 Artificial Intelligence and Capsule Endoscopy: Automatic Detection of Gastric Vascular Lesions Using a Convolutional Neural Network

Abstract

Introduction: Capsule endoscopy (CE) is the main diagnostic modality for the investigation of obscure gastrointestinal bleeding. Although it is designed to identify small bowel pathology, lesions (particularly vascular lesions) are frequently documented in locations within reach of conventional endoscopy. However, the performance of CE for detection of gastric lesions has been shown to be suboptimal. Studies on the development of artificial intelligence (AI) systems, particularly convolutional Neural Networks (CNN), for automatic analysis of CE images have provided promising results. To date, performance of these automated algorithms for detection of gastric vascular lesions in CE images has not been evaluated. We aimed to develop and test a CNN-based algorithm for automatic detection of gastric vascular lesions (angiectasia, gastric antral vascular ectasia, varices, and red spots). Methods: A total of 1275 CE images were included for construction of the CNN, 470 containing vascular lesions and 805 showing normal mucosa. For automatic detection of gastric lesions, these images were inserted into a CNN model with transfer learning. A training dataset comprising 80% of the total pool of images was defined. Subsequently, we evaluated the performance of the network using an independent test dataset (20% of total image pool). The output provided by the CNN was compared to a consensus classification provided by two endoscopists experienced in CE (more than 1000 CE videos each). We calculated the sensitivity, specificity, accuracy, positive predictive and negative predictive values (PPV and NPV, respectively), and area under the curve (AUC). Results: After optimization of the neural architecture of the algorithm, our model was able to detect detected gastric vascular lesions with a sensitivity, specificity, PPV and NPV of 86.2%, 98.1%, 96.4% and 92.4%, respectively. The algorithm had an overall accuracy of 93.7%. The AUC was 0.97. The CNN read the validation dataset in 6 seconds (average processing speed of 45 images per second). Conclusion: This is the first report of a deep learning system for automatic detection of gastric vascular lesions in CE images. The implementation of these systems may potentiate the ability of CE for exploration of suspected upper gastrointestinal disease, which may be particularly helpful in patients with limited tolerance to EGD.