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

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Photoacoustic computed tomography (PACT) using linear transducer arrays is cost-effective but suffers from poor 3D elevational resolution.
• Anisotropic spatial resolution limits the diagnostic accuracy of volumetric PACT.

Purpose of the Study:

• To introduce 3D multiview FISTA-Net (MV-FISTA-Net), a deep learning framework to enhance elevational resolution in linear-array PACT.
• To improve the overall 3D spatial resolution and reduce anisotropy in PACT imaging.

Main Methods:

• Developed a model-based deep learning framework by unrolling the Fast Iterative Shrinkage-Thresholding Algorithm (FISTA) into a 3D architecture.
• Integrated multi-view elevational scan data into the MV-FISTA-Net model.
• Employed a two-stage training strategy: pretraining on synthetic data and fine-tuning on a mix of synthetic and experimental data.

Main Results:

• MV-FISTA-Net demonstrated up to a 1.72x improvement in elevational resolution compared to deconvolution-based methods.
• Achieved up to an 8.57x resolution improvement compared to single-scan PACT.
• Showcased efficiency with over 20x faster inference time than MV-FISTA and maintained resolution advantage with fewer views.

Conclusions:

• MV-FISTA-Net effectively enhances elevational resolution and reduces anisotropy in volumetric linear-array PACT.
• The proposed deep learning framework is a generalizable and practical solution for improved 3D PACT imaging.
• MV-FISTA-Net offers significant advancements in resolution and computational efficiency for PACT applications.