Building an Automated Module for Image Quality Assessment from Narrow-Banding-Imaging Endoscopy Cameras

Abstract

Gastrointestinal diseases have a significant impact on human well-being, and among them, gastrointestinal tumors have particularly high rates of occurrence and death. Besides traditional endoscopy, with the advancements in technology, endoscopic techniques nowadays such as the Narrow-banding-imaging (NBI) technique play a fundamental role in the quality of life for individuals by providing numerous imaging information for accurate diagnosis of gastrointestinal diseases. Although the amount of imaging data is abundant, the majority of them are substandard, making it extremely difficult for physicians to find quality images, therefore it is necessary to build an image quality control module for data cleaning. Based on the characteristics of the gastroscopy environment and inspired by non-reference image quality assessment (NR-IQA) methods, in this work, we utilized an IQA framework that assesses image quality from NBI endoscopy cameras that composes two stages utilizing deep learning approaches with some improvements. The first stage is based on a patch-based classification model to assess the quality of multiple regions of the image, which uses local features extracted from multi-layer of the convolutional neural network (CNN). In the second stage, these results of patches will be aggregated to give the final quality level for the entire image. Our improved pipeline shows over 96% and 97% with respect to overall precision and recall respectively. In addition, the final quality level of the endoscopic images was clinically evaluated by medical experts from Viet Duc and K Tan Trieu hospital. The results show that output quality levels are highly correlated to medical professionals’ perception of endoscopic image quality. In terms of storage, the IQA module not only reduces nearly 90% the amount of storage capacity needed but also provides users a lot of flexibility depending on different scenarios. Finally, with the inference speed improvement method, we achieved 48 FPS in terms of frame rates, which is 4 times faster than the original.