Attention-driven complementary information fusion network for sparse photoacoustic image reconstruction.

Photoacoustic tomography (PAT) is an emerging biomedical imaging modality that uniquely combines high spatial resolution with deep tissue penetration in a non-invasive manner, holding significant promise for diverse applications. However, image reconstruction quality in PAT severely degrades under limited-view data acquisition scenarios, such as those imposed by the physical constraints of intracavitary imaging. Conventional reconstruction methods (e.g., Delay-and-Sum, DAS) under these conditions typically yield images plagued by severe artifacts and loss of fine structural details. While deep learning (DL) approaches offer some improvement, existing post-processing methods still struggle to accurately recover intricate anatomical features from severely undersampled, limited-view data, often resulting in blurred details or persistent artifacts. To address these critical limitations, we propose DUAFF-Net, a novel dual-stream deep learning architecture. DUAFF-Net uniquely processes two complementary input representations in parallel: 1) conventional DAS reconstructions, and 2) pixel-wise interpolated raw data. The network employs a sophisticated two-stage feature fusion strategy to maximize information extraction and synergy. In the first stage, the Multi-scale Information Aggregation and Feature-refinement Module (MIAF-Module) enables early-stage cross-modal information complementarity and feature enhancement. Subsequently, the Global Context and Deep Fusion Module (GCDF-Module) focuses on holistic feature optimization and deep integration across the streams. These modules work synergistically to progressively refine the reconstruction. Extensive experiments on simulated PAT datasets of retinal vasculature and complex brain structures, as well as an