Automated Gastrointestinal Disease Recognition for Endoscopic Images

The human Gastrointestinal (GI) tract can be affected by different diseases and endoscopy has been seen to perform well for diagnosing GI tract problems. Accurate identification of underlying problems in GI tract endoscopic images is important as it affects decision-making on treatment and follow-up. In developing countries trained endoscopic experts are small in number and expensive. Even though medical recognition is a promising field of application for Artificial Intelligence (AI) publicly available datasets for such tasks are small in number. Kvasir dataset is one of the publicly available medical datasets. It consists of gastrointestinal endoscopic images that belong to eight different classes. We have automated recognition of GI tract landmarks and diseases, for classes that are available in Kvasir, with the use of Convolutional Neural Networks (CNNs). CNNs are widely used for visual recognition due to their ability to capture local features and their computational efficiency compared to fully connected networks. We have fine-tuned a residual model based on ResNet50 and a dense model based on DenseNet121 on Kvasir dataset. The models’ performance on a test set that consists of 75 images from each class is 86.9% for dense model and 87.8% for residual model. We have also built a user interface for users to select images and get recognition results. The interface built can serve as a decision support system for classifying GI tract endoscopic images. It can also further be extended for recognition in videos by feeding the video input as a sequence of images.