Fine-Grained Gastrointestinal Endoscopy Image Categorization

Abstract

Gastrointestinal endoscopy is of great significance to improve the accuracy and efficiency for diagnosis of digestive tract diseases. With the development of artificial intelligence in medical images, the computer-assisted system of diagnosis is developed to assist specialists in gastrointestinal endoscopy diagnosis. Convolutional Neural Networks (CNNs) are good at recognizing significant categories differences, but poor at subtle inter-class differences. The images captured in gastrointestinal endoscopy have subtle inter-class differences among sub-categories, so fine-grained gastrointestinal endoscopy image classification is more difficult than ordinary image classification tasks. To address this challenge, this paper used Recurrent Attention Convolutional Neural Network (RACNN) to transfer learning image's features and label smoothing regularization method to improve experimental performance. Experimental results show that the RACNN with label smoothing technique achieves the best classification performance of traditional deep neural networks.