G-NET LIGHT AND RU-NET ASSISTED GOOGLENET V4 FRAMEWORK FOR EXPLAINABLE GASTROINTESTINAL ENDOSCOPY IMAGE ANALYSIS

Abstract

Gastrointestinal (GI) endoscopy is a very common test used for early detection of the diseases like polyps, ulcer, bleeding and inflammatory conditions. However, while it can be considered a useful tool for diagnosing, the inspection of many endoscopes is a time-consuming process and needs experienced physicians, which means there could be some differences in diagnosing. To solve this problem, we introduce a deep learning method of automatic segmentation and classification of GI endoscopy images.The endoscopy images were processed using the following techniques, such as Contrast Limited Adaptive Histogram Equalization (CLAHE), removal of the noise, normalizing the intensity, and adjusting the size of the image. Then, G-Net Light was used to detect informative features, and RU-Net was applied for lesion segmentation. The segmentation algorithm obtained a Dice coefficient of 72.6%, IoU of 60.8%, and pixel accuracy of 91.8% which is good enough to segment lesions.According to the experiment, GoogleNet V4 performed better than GoogleNet V2 concerning overall accuracy, precision, recall, specificity, F1-score, and AUC with values of 93.33%, 92.84%, 93.10%, 94.21%, 92.97% and 0.965 respectively. That is, GoogleNet V2 obtained 90.42% accuracy and AUC of 0.934.