Abstract:
limitations with instrumentation and dataacquisition. Common challenges include having a limited number of available acoustic detectors and a reduced “view” of the imaging target. Forming an image with classical reconstruction methods from insufficient data often result in images with artifacts that degrade image quality. Advanced methods such as iterative reconstruction can be effective in reducing or removing the artifacts. But these methods are also computationally expensive and might not be appropriate in settings requiring near realtime imaging. In this work, we summarize our efforts in utilizing deep learning to address xiv the deficiencies of sparse spatial sampling and limited-view detection in PAT image reconstruction. We begin with an introduction to fundamental principles of photoacoustic imaging (Chapter 1). This is followed by a brief introduction to deep learning and summarize commonly used deep learning frameworks for PAT image reconstruction (Chapter 2). Next, we describe a novel convolutional neural network architecture termed Fully Dense UNet for sparse PAT image reconstruction (Chapter 3). We then describe pixel-wise deep learning, a data pre-processing step that seeks to provide a more informative input to the neural network (Chapter 4). Next, we describe a modified network architecture termed Dense Dilated UNet that leverages the benefits of dense connectivity and dilated convolutions for 3D PAT image reconstruction (Chapter 5). We then describe Fourier Neural Networks as a fast and general solver for the photoacoustic wave equation (Chapter 6). Finally, we conclude with a discussion of key challenges in using deep learning for PAT image reconstruction and future work (Chapter 7).
