Compensating for the spatial impulse response in 3D photoacoustic computed tomography using a learned data space restoration method

Photoacoustic computed tomography (PACT) is a promising modality for biomedical imaging. Many existing reconstruction algorithms for photoacoustic computed tomography (PACT) assume that the ultrasound transducers deployed to record the pressure measurements are point-like. However, using transducers with large detecting areas can be advantageous due to enhanced sensitivity and signal-to-noise ratio. Such transducers with finite detection surfaces have anisotropic detection sensitivity described by the spatial impulse response (SIR). When not accounted for, the reconstructed images may suffer from severe, spatially varying degradation in spatial resolution. This study proposes a learning-based method to compensate for SIR in the measurement domain, mapping the pressure data acquired by an array of large rectangular transducers to idealized data corresponding to point-like transducers. The model was trained with analytically computed pressure data from uniform spheres with stochastically determined locations, sizes, and initial pressure values. Subsequently, it was tested on pressure data simulated from realistic numerical breast phantoms. Importantly, although trained on simplified objects, the proposed method demonstrates robust performance on numerical breast phantoms that contain complex anatomical structures unseen in training.