Classification of upper gastrointestinal tract diseases using endoscopic images

Abstract

Automatic classification and disease detection in medical images, aided by machine learning, provide crucial support to prevent overlooked instances and ensure prompt treatment of diseases. Despite impressive achievements in the field of polyp detection from endoscopic images, classification of other diseases, such as reflux esophagitis, esophageal cancer, gastritis, gastric cancer, and duodenal ulcer, is still subject to significant limitations and remains a challenging area of study because of their different and more challenging characteristics. This paper proposes a method to roughly classify the diseases from the whole images by deep learning. In particular, we focus on identifying hard samples from the training dataset and enriching them with some fundamental augmentation techniques. We then employ a cutting-edge model, specifically ResNet, for the final classification stage. Additionally, we enhance the original ResNet’s loss function by incorporating another loss function called focal loss. These modifications play a crucial role in boosting the accuracy of the ResNet model. Our proposed method outputs the disease category and corresponding heat map showing the area of interest. It achieved very promising accuracy (99.55%) for the classification of five lesions on our self-collected dataset. It serves a dual purpose. Firstly, it aids in the training of novice endoscopists, enabling them to gain valuable experience. Secondly, it offers a rapid solution for annotating extensive volumes of endoscopic image data at the label level.