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

Area of Science:

• Biomedical Imaging
• Medical Physics
• Computer Vision

Background:

• Photoacoustic tomography (PAT) offers high-resolution, deep-tissue imaging but suffers from image degradation with limited-view data.
• Conventional methods like Delay-and-Sum (DAS) produce artifacts and lose detail in limited-view scenarios.
• Existing deep learning (DL) methods struggle to fully recover intricate details from severely undersampled limited-view PAT data.

Purpose of the Study:

• To develop a novel deep learning architecture, DUAFF-Net, for high-quality image reconstruction in limited-view photoacoustic tomography.
• To address the limitations of conventional and existing DL methods in recovering fine anatomical features from incomplete data.
• To enhance the robustness and performance of PAT imaging under challenging data acquisition constraints.

Main Methods:

• Proposed DUAFF-Net, a dual-stream deep learning network processing conventional DAS reconstructions and interpolated raw data.
• Implemented a two-stage feature fusion strategy: Multi-scale Information Aggregation and Feature-refinement Module (MIAF-Module) and Global Context and Deep Fusion Module (GCDF-Module).
• Validated DUAFF-Net on simulated retinal vasculature, brain structures, and in vivo mouse abdomen datasets.

Main Results:

• DUAFF-Net generated high-quality images from highly incomplete limited-view PAT data.
• Achieved significant improvements over DAS: ~18.38 dB PSNR and ~0.69 SSIM.
• Outperformed state-of-the-art DL reconstruction models in preserving fine details and suppressing artifacts.

Conclusions:

• DUAFF-Net demonstrates superior performance for limited-view PAT reconstruction, even with severely incomplete data.
• The dual-stream architecture and feature fusion strategy effectively enhance image quality and detail recovery.
• DUAFF-Net offers a robust solution for challenging PAT applications where data acquisition is constrained.