Acoustic resolution photoacoustic microscopy imaging enhancement: learning based and model based approaches

• Main Author: Zhang, Zhengyuan
• Other Authors: Zheng Yuanjin
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Acoustic resolution photoacoustic microscopy; Neural network; Photoacoustic imaging; Model based algorithm
• Online Access: https://hdl.handle.net/10356/177423
• Institution: Nanyang Technological University

Photoacoustic Imaging, a novel medical imaging modality, holds great promise for non-invasive bio-tissue imaging and widespread clinical applications. Its unique combination of optical imaging contrast and acoustic imaging penetration depth delivers outstanding performance surpassing individual modalities. Operating on the principle of the photoacoustic effect, a pulsed laser irradiates the imaging sample, generating acoustic waves via thermo-elastic expansion, which are then detected using transducers and reconstructed using ultrasonic imaging algorithms. Two main implementations exist within photoacoustic microscopy: Optical Resolution Photoacoustic Microscopy (OR-PAM) and Acoustic Resolution Photoacoustic Microscopy (AR-PAM). While the latter achieves deep penetration, its resolution and quality lag behind the former. This prompts investigation into enhancing AR-PAM imaging quality while retaining its depth and speed advantages through computational algorithms. The enhancement focused on contrast/resolution improvement, noise reduction, and imaging speed acceleration. To address the above challenge, two algorithmic perspectives emerge: learning-based and model-based approaches. Deep learning algorithms have been proposed as first solution, leveraging synthetic training data to reconstruct PAM images, enhancing resolution and reducing noise. In this thesis, MultiResU-Net incorporating Multi-Res blocks and Res-Path modules has been proposed, targeting vasculature enhancement in AR-PAM images. Domain adaptation technique under the GAN framework further refines image quality. Subsequently, a dual-branch fusion network proposes a hybrid enhancement framework, combining software and hardware for simultaneous resolution and speed enhancement, resulting in HSD-PAM approach with significant improvements. Model-based algorithms formulate an optimization problem from a Bayesian perspective, which involves data-fidelity and regularization terms. Variable splitting techniques, like the HQS method, can be employed to solve this problem, iterative schemes are developed to enhance the AR-PAM image with deep PnP prior. What’s more, a novel GSDP-PAM approach integrates group sparsity and deep PnP priors is proposed, which further enhances AR-PAM imaging quality. Experimental validation involves enhancing simulated and in vivo AR-PAM images, assessing effectiveness through quantitative metrics like PSNR, SSIM, CNR, and FWHM. Application to in vivo AR-PAM images demonstrates enhancement of imaging quality through software algorithms, eliminating the need for renovating bulky physical setups. Comparative studies highlight the superiority of proposed algorithms, with enhanced resolution and narrowed gap between AR-PAM and OR-PAM. The results suggest potential deployment of these methods in real AR-PAM systems, enabling high-resolution, deep penetration imaging at high imaging speeds for clinical applications.