Over the past decade, photoacoustic imaging (PAI), a novel hybrid imaging modality, has gained a lot of attention in medical imaging due to its ability to render high-resolution images with good optical contrast. In particular, photoacoustic tomography/photoacoustic computed tomography (PAT/PACT) is the widely preferred embodiment of PAI for deep tissue imaging. Conventionally in PAT systems, nanosecond pulsed lasers are used as excitation sources to irradiate the sample homogeneously. The generated PA waves are recorded around the boundary of the sample using either single-element ultrasound detectors or ultrasound detector arrays. The acquired PA waves are then reconstructed into crosssectional PAT images representing the structural and functional image of the chromophores. Various reconstruction algorithms can be used, such as simple delay-and-sum beamformer, frequency domain reconstruction, iterative reconstruction, model-based reconstruction, etc. Despite the numerous advantages of PAT imaging systems, their potential clinical adaptation is hindered by limitations such as resolution degradation, image distortion, slow imaging speed, etc. Moreover, attempts to reduce their cost by using multiple single-element detectors instead of array detectors and cheaper laser sources such as pulse laser diodes further impair image quality.

In recent years, deep learning, a subset of machine learning, has gained significant interest due to its ability to solve the limitations associated with imaging-related tasks. In particular, convolutional neural networks (CNNs) are most widely preferred for such applications. In this Spotlight, we will introduce how to employ deep learning techniques to overcome some of the limitations in the circular geometry-based PAT systems. We begin with an introduction to PAT imaging and discuss the challenges associated with it. This is followed by a hands-on introduction to deep learning with an emphasis on CNN, and how to adapt it to tackle the limitations in circular geometry-based PAT systems. We also elaborately discuss three examples of how to apply deep learning to overcome and improve some real-world problems encountered in circular geometry-based PAT systems. Overall, this spotlight will provide readers with an understanding of deep learning basics and how to adapt them for various applications in PAT imaging systems.

We are grateful to the editors of the SPIE Spotlight series. Without their support, this Spotlight would not have been possible. We thank the referees for their valuable comments. We thank Malik for proofreading the manuscript.