A Learnable Physical Deep Learning Model for Sparse Limited View Photoacoustic Tomography.

Area of Science:

• Biomedical Imaging
• Medical Physics
• Artificial Intelligence

Background:

• Photoacoustic tomography (PAT) systems typically require extensive transducer arrays for high-quality clinical imaging.
• Manufacturing these arrays is costly and limits clinical applications.
• Sparse reconstruction algorithms can potentially reduce hardware demands.

Purpose of the Study:

• To develop a novel deep learning model for photoacoustic image reconstruction using sparse and limited-view data.
• To alleviate the manufacturing cost and complexity associated with extensive photoacoustic transducer arrays.

Main Methods:

• A learnable physical deep learning model was designed for photoacoustic image reconstruction.
• The model processes signals acquired by sparse photoacoustic array elements.
• Performance was evaluated using numerical simulations and in vivo experiments.

Main Results:

• The deep learning model achieved high-quality, artifact-free photoacoustic tomography reconstruction from sparse signals.
• Even with extremely sparse data (16 detectors, 155 degrees), significant results were obtained.
• Reconstruction yielded a structural similarity of 0.728 ± 0.029 and a peak signal-to-noise ratio of 20.700 ± 2.404 dB.

Conclusions:

• The proposed learnable physical deep learning model effectively reconstructs photoacoustic images from sparse data.
• This approach offers a cost-effective and flexible solution for clinical photoacoustic tomography.
• The findings demonstrate the potential of AI in advancing sparse-view photoacoustic imaging.