Three-Dimensional Siamese Multi-Level Features Neural Network Based 3D Fusion Improves the Depth of Field in Photoacoustic Microscopy.

Abstract

Microscopic imaging techniques pursue high-resolution, large depth of field (DoF) imaging but are limited by hardware, especially the strong focusing of objective lenses. Optical-resolution photoacoustic microscopy (OR-PAM) has a narrow DoF due to the intense laser focusing needed for high-resolution imaging. To address this, we propose a novel volumetric information fusion method using a three-dimensional siamese multi-level features convolutional neural network (3DSMFCNN) for cost-effective, large-DoF imaging. Initially, an initial decision map (IDM) is produced by performing focus region identification on multi-focus 3D photoacoustic data with the pre-trained 3DSMFCNN. The IDM is then refined through consistency verification and Gaussian filtering to generate the final decision map (FDM). A DoF-enhanced photoacoustic image is obtained by voxel-weighted averaging based on the FDM. Experiments with multi-focus 3D simulated fibers, blood vessels, and real data demonstrate that the method significantly extends the DoF of OR-PAM without sacrificing lateral resolution, which confirms its effectiveness, robustness, and applicability.