Automatic Classifying Informative Frames for Upper Gastrointestinal Endoscopy Diagnosis Using Deep Learning Method
摘要
Deep learning models play important roles to support endoscopists, especially in detecting and segmenting abnormalities in the gastrointestinal tract. However, when using endoscopy image sequences or videos as the diagnosing mean, there are limitations in terms of visual quality such as blurring and glaring. Those images are called non-informative which degrades significantly the performances of the deep learning models. In this work, we propose a robust deep learning-based solution to filter out non-informative images in four different endoscopy lighting modes. The proposed method is tested in a practical dataset. The experimental results show that precision is 90.29%, Recall is 91.51%, and F1-score is 90.79%. Furthermore, for the subjective evaluations with participants of endoscopy experts confirm that the proposed model is a feasible solution in their daily clinic. They comment that the proposed method is non-biased.