Deep Learning Applications in Light-based Ultrasound Imaging: Photoacoustic Microscopy, Photoacoustic Tomography, and Passive Cavitation Mapping

Abstract

Medical imaging systems inevitably encounter fundamental physical limitations that have traditionally been addressed either through costly, time-intensive hardware innovation or by accepting tradeoffs in resolution, image quality, imaging speed, or overall system usability and commercialization potential. The emergence of advanced computational methods, particularly deep learning, offers a fundamentally different paradigm: rather than working around hardware constraints, these limitations can be mitigated through software. This shifts the primary challenge in imaging development from physical tradeoffs to problems of dataset generation, model design, optimization, and inference robustness. In this thesis, we present deep learning frameworks that address long-standing physical constraints in light-based ultrasound imaging, specifically photoacoustic microscopy (PAM), photoacoustic computed tomography (PACT), and passive cavitation mapping (PCM). In PAM, we investigate the limitation imposed by laser repetition rates and the resulting tradeoff between spatial sampling density and imaging speed for a fixed field of view. We develop a deep learning-based upsampling approach that restores high-quality images from sparsely sampled data without sacrificing acquisition speed. In PACT, we demonstrate that conventional resolution limits can be surpassed by coupling a digital model of a 512-element ring-array transducer with supervised deep learning deconvolution. In PCM, we use deep learning to characterize the complex interactions of cavitation collapses during kidney stone lithotripsy and generate predictive visualizations of stone material loss, enabling improved treatment monitoring and effectiveness.