Enhancing Volumetric Imaging in Linear-Array Photoacoustic Tomography: multiview fusion with deep learning

Photoacoustic computed tomography (PACT) with linear transducer array is a widely adopted imaging modality due to its cost-effectiveness and portability, yet its three-dimensional (3D) imaging suffers from severe anisotropic spatial resolution, with the elevational resolution being substantially worse than the axial and lateral resolutions. To address this limitation, we propose 3D multiview FISTA-Net (MV-FISTA-Net), a model-based deep learning (DL) framework that unrolls the Fast Iterative Shrinkage-Thresholding Algorithm (FISTA) into a learnable 3D architecture and integrates information from elevational scans taken from multiple views. To bridge the gap between synthetic numerical data and real experimental acquisitions, we design a two-stage training strategy: Stage I pretrains the network on paired synthetic data, while Stage II fine-tunes on a balanced mixture of synthetic and experimental measurements. We validate the method on simulation, phantom, and in vivo human datasets. Experimental results show that MV-FISTA-Net achieves enhanced elevational resolution compared to deconvolution-based MV-FISTA, with up to a 1.72× improvement. Compared to a single scan, MV-FISTA-Net achieves up to a 8.57× resolution improvement. Notably, even with data from only 5 scanning views, MV-FISTA-Net maintains its advantage in resolution over 9-view MV-FISTA, underscoring its efficiency and robustness. MV-FISTA-Net also achieves more than a 20× improvement in inference time compared to MV-FISTA. These results highlight MV-FISTA-Net as a generalizable and practical solution for volumetric linear-array PACT with reduced anisotropy.