Fine-tuning Pre-trained Convolutional Neural Networks for Gastric Precancerous Disease Classification on Magnification Narrow-band Imaging Images

Abstract Gastric cancer(GC) is the fourth leading cause of cancer death worldwide. To prevent the occurrence of advanced GCs, there is a need for immediate detection and treatment of gastric precancerous and early cancerous lesions. Magnification endoscopy with narrow-band imaging (M-NBI) system as an advanced diagnostic imaging technology is widely used in evaluating gastric lesion types, which can interpret gastric lesion characteristics by enhancing contrasts between vessels and mucosal surfaces. Based on microvascular morphologies presented on M-NBI images, physicians can manually diagnose gastric lesions; but this is a tough work for unexperienced doctors and it is lacking of objectivity. In this study, we propose a transfer learning framework by fine-tuning pre-trained convolutional neural networks (CNNs) to classify gastric M-NBI images into three classes: chronic gastritis (CGT), low grade neoplasia (LGN) and early gastric cancer (EGC). The method we choose is used to compare with three kinds of traditional handcraft texture feature extraction methods and CNN models trained directly by our dataset. Results show that the performance of fine-tuned CNNs outperforms traditional handcraft features and trained CNNs. Experiments also illustrate that ResNet50 can achieve 0.96 accuracy, 0.92, 0.91 and 0.99 f1-scores for classifying M-NBI images into CGT, LGN and EGC. In conclusion, the proposed framework is suit for multi-classification tasks of gastric M-NBI images.