Abstract
Photoacoustic tomography (PAT), widely applied using linear array ultrasound transducers for clinical and preclinical imaging, faces significant challenges due to sparse sensor arrangements and limited sensor pitch. These factors often compromise image quality, particularly in devices designed to have fewer sensors to reduce complexity and power consumption, such as wearable systems. Conventional reconstruction methods, including delay-and-sum and iterative model-based techniques, either lack accuracy or are computationally intensive. Recent advancements in deep learning offer promising improvements. In particular, model-based deep learning combines physics-informed priors with neural networks to enhance reconstruction quality and reduce computational demands. However, model matrix inversion during adjoint transformations presents computational challenges in model-based deep learning. To address the challenges, we introduce a simplified, efficient GE-CNN framework specifically tailored for linear array transducers. Our lightweight GE-CNN architecture significantly reduces computational demand, achieving a 4-fold reduction in model matrix size (2.09 GB for 32 elements vs. 8.38 GB for 128 elements) and accelerating processing by approximately 46.3 %, reducing the processing time from 7.88 seconds to 4.23 seconds. We rigorously evaluated this approach using synthetic models, experimental phantoms, and in-vivo rat liver imaging, highlighting the improved reconstruction performance with minimal hardware.