Develop AI based image restoration algorithm for deep tissue imaging in photoacoustic system

• Main Author: Hao, Zeliang
• Other Authors: Zheng Yuanjin
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Engineering
• Online Access: https://hdl.handle.net/10356/179288
• Institution: Nanyang Technological University

In today's global landscape, information technology has become ubiquitous. Over the years, it has undergone constant transformation and is poised to continue evolving. From the inception of computers to the advancements in software and artificial intelligence, a new epoch has unfolded, significantly impacting human life. Deblurring images poses a fundamental challenge within the realm of low-level computer vision, and the objective is the restoration of a clear image from a blurred input. The advent of deep learning has brought in a significant improvement in tackling this challenge, giving rise to a multitude of deblurring networks. In this paper, the author categorizes the image restoration algorithms for deep tissue imaging in photoacoustic systems as a generalized image deblurring problem and conduct a series of studies based on it. This study proposes a new network of image restoration algorithms at to overcome the problems encountered in photoacoustic imaging. Photoacoustic imaging(optoacoustic imaging), is a biomedical imaging technique rooted in the photoacoustic effect. The Photoacoustic imaging results are degraded due to inherent imaging mechanism, imperfection of imaging equipment and random additive white gaussian noise, which result in blurring, noising and even heavy artifacts presented in the raw image. This project propose a novel approach that integrates two GAN models, namely Blurring GAN (GAN-Blur) and Deblurring GAN (GAN-Deblur), to learn a superior image deblurring model by primarily mastering the art of image blurring. The initial model, GAN-Blur, is trained to induce blurring in sharp images through an unpaired dataset containing both sharp and blurred images, guiding the second model (GAN-Deblur) in learning the proper deblurring of such images. To minimize the disparity between real and synthetic blurring, the author use the combination of three loss. Additionally, to better model the blurring of photoacoustic imaging systems, the author also uses a Real-World Blurred Image (RWBI) dataset, which contains lots of real blurry images collected by different equipment. The experimental results illustrate that the proposed approach consistently surpasses existing methods in terms of quantitative metrics on both the RWBI dataset and the commonly utilized public GOPRO dataset. The author conducted an in-depth analysis of the obtained results and provided possible explanations. Furthermore, this study has presented suggestions for future research endeavors aimed at tackling areas that could be enhanced.