Development of Low Complexity Encoder and Summarization Techniques for Wireless Capsule Endoscopy Video

Abstract

Wireless capsule endoscopy (WCE) is the state-of-the-art medical procedure for scanning the entire digestive tract to diagnose gastrointestinal (GI) diseases. Its non- invasiveness and ease of usage make it a better option than conventional endoscopy. However, it is inferior to conventional endoscopy due to low image quality imposed by capsule’s complexity and power consumption. In one complete scan of GI tract, a capsule captures between 90000 and 180000 frames during its peristalsis movement. Diagnosing such a large number of images is a time-consuming and tedious procedure that needs a gastroenterologist’s undivided attention. The main aim of the research work is two folds. One involves the development of a low complexity video encoder that can reduce the computations in the capsule. The other part involves a WCE video summarization framework to provide an efficient diagnosis. Developing a low-complexity video encoding architecture that can achieve high compression performance at a low bit rate while maintaining acceptable reconstruction quality is a challenging task in WCE. A distributed video coding (DVC) architecture is proposed to achieve this, which transfers encoder complexity to the decoder side. It employs a keyframe encoder that takes advantage of GI image textural properties to reduce the complexity. Furthermore, the low-frequency bands of the Wyner-Ziv (WZ) frames are used as auxiliary information at the decoder to generate high-quality side information that enables the encoding of high frequency bands with a low bit rate. The proposed DVC framework is further enhanced to reduce complexity by eliminating WZ-chroma component encoding. Exploiting the similarity in colour and texture properties between consecutive frames in WCE video, a deep convolutional neural network model is integrated into the decoder side to predict the chroma component. The proposed methods achieve improvements in coding gain with low complexity encoder when compared with benchmark compression schemes. A physician must dedicate lot of time in reviewing the large number of frames, and there is a considerable risk of missing frames that are associated with lesion symptoms. Review time can be minimized by extracting the summary of WCE video by eliminating the redundant frames. To achieve this, a summarization framework consisting a shot boundary detection and keyframe extraction methods is presented.The proposed framework achieves better summarization performance measured us- ing F-score and compression ratio compared to state of the art WCE summarization methods.