OVERCOMING OPTICAL SCATTERING IN PHOTOACOUSTIC IMAGING WITH INTENSITY-RECOVERING DEEP LEARNING MODEL

Wu, Christine Szu-Yao

Permalink

https://hdl.handle.net/2142/124906 Copy

Description

Title

OVERCOMING OPTICAL SCATTERING IN PHOTOACOUSTIC IMAGING WITH INTENSITY-RECOVERING DEEP LEARNING MODEL

Author(s)

Wu, Christine Szu-Yao

Issue Date

2021-05-01

Keyword(s)

Photoacoustic Imaging; Optical Scattering; Convolutional Neural Network; Conditional Generative Adversarial Network; Image-to-image Translation

Date of Ingest

2024-10-16T09:34:51-05:00

Abstract

Photoacoustic imaging (PAI) is a hybrid imaging modality with rich optical contrast and high spatiotemporal resolution. PAI utilizes the principles of electromagnetic energy absorption and thermal expansion of different specimens as a contrast to generate ultrasound waves and to visualize deeper biological structures than pure optical imaging modalities find difficult to image in a non-invasive manner. However, optical scattering and ultrasound attenuation in biological tissues deteriorates the quality of PAI. The penetration depth is still limited to several optical mean-free-path, disabling current PAI techniques from numerous potential clinical applications, especially on humans. While many reconstruction algorithms improved contrast to noise ratio, the issue with scattering in optical imaging remains. In recent years, deep learning methods have been infused into various imaging applications and have demonstrated promising results in the context of medical imaging. In this thesis, we first constructed a laser intensity predictive network NetP, which is based on a convolutional neural network (CNN), to predict the laser intensity given a PA image. Then, we integrated NetP with a general solution image-to-image translation conditional Generative Adversarial Network (cGAN) to construct an intensity-recovering PowerNet. PowerNet takes in an additional layer of laser intensity on top of one ultrasound and one photoacoustic image to assist the cGAN generator in overcoming laser attenuation in current PA images. We also incorporated SSIM score and laser intensity difference into the loss calculation of PowerNet for a more robust learning evaluation while performing image-to-image translation between ultrasound and photoacoustic images. We trained the PowerNet on lab-generated ideal datasets and real arm blood vessel datasets to evaluate its performance and its practicality in realworld clinical applications. The resulting ultrasound-assisted PA images from our method can retain uniform laser intensity while depth increases within the tissue. The PowerNet consistently generates PA images with minimal intensity attenuation as compared to the state-of-the-art methods. Thus, our method has effectively reduced the amount of optical scattering in PAI.

Type of Resource

text

Genre of Resource

dissertation/thesis

Language

eng

Owning Collections