Real-time gastric intestinal metaplasia semantic segmentation with multiple abnormalities using deep learning approach

Abstract

This thesis declares the segmentation of gastric intestinal metaplasia (GIM) in real-time. Recently, GIM segmentation of endoscopic images has been conducted to distinguish GIM from a healthy stomach. However, achieving real-time detection is difficult. Challenging conditions include multiple color modes (white light endoscopy and narrow-band imaging), other abnormal lesions (erosion and ulcer), noisy labels, etc. Herein, our model is based on BiSeNet and can overcome the many issues regarding GIM. Applying auxiliary head and loss boosts the performance on multiple color modes. In addition, pre-processing techniques, including location-wise negative sampling, jigsaw augmentation, and label smoothing, are utilized to improve detection performance. Finally, the decision threshold can be independently altered for each color mode. Work undertaken at King Chulalongkorn Memorial Hospital examined 940 histologically proven GIM images and 1239 non-GIM images, obtained over 173 FPS. In terms of accuracy, our model outperforms all baselines. Our results reveal F1-score, sensitivity, specificity, accuracy, and mean intersection over union (IoU), achieving 91%, 91%, 96%, 94%, and 55%, respectively. In addition, the effectiveness of the proposed methods was validated on baseline models, achieving F1-score and IoU values of 93% and 56% for STDC2-Seg50 and 93% and 56% for BlazeNeo.