Non-Local Attention and Densely-Connected Convolutional Neural Networks for Malignancy Suspiciousness Classification of Gastric Ulcer

Abstract

Gastric ulcer is one of the most common types of stomach disease. Malignancy suspiciousness classification of gastric ulcer is a crucial indicator for early cancer detection and prognosis. Technically, this problem suffers from the complexity and variability of endoscopic pathological images. For addressing these challenges, we propose a deep learning based classification neural network which combines the densely-connected architecture and non-local attention mechanism. Structurally, we add the attention block into the cascaded dense blocks for catching more contextual information and enhancing the correlation between pixels and regions. Experimentally, we implement sufficient experiments on our own gastroscopic image dataset, which is delicately annotated twice per image by medical specialists. Quantitative comparisons against several prior state-of-the-art methods demonstrate the superiority of our approach. As a result, we achieve an overall diagnostic accuracy of 96.79 %, a recall of 94.92% and an F1-score of 94.70 %, close to the diagnostic level of a gastroenterologist. The area under the receiver operating characteristic (ROC) curves of the deep learning model achieve an average of 0.93.