Deep Learning for scaling large-aperture photoacoustic computed tomography : From single fingers to the human hand

Photoacoustic Computed Tomography (PACT) leverages the photoacoustic effect for high-resolution anatomical and molecular imaging. We developed an advanced PACT system using eight conventional linear arrays arranged in a half-ring geometry, achieving a balance between cost-efficiency and enhanced image quality through a large-aperture detection setup. Although this large-aperture PACT system provides high-quality imaging for in-vivo human applications, it is susceptible to optical shadowing and misalignment issues between optical paths and detection planes, particularly during complex and large-target imaging, such as imaging of the human hand. These issues can lead to degraded image quality. To address these issues, we implemented an encoder-decoder structure-based deep learning (DL) enhancement strategy. The DL model was initially trained using a paired PACT single-finger dataset, which included images obtained with full detection using all eight transducers and those with low detection using fewer transducers or elements. For human-hand PACT imaging, the DL-enhanced system, trained exclusively with the single-finger dataset, effectively mitigated the image quality issues by improving contrast-to-noise ratios and the clarity of vessel structures. These findings validate the efficacy of the DL-enhanced PACT system for complex anatomical imaging applications, such as diagnosing peripheral arterial disease.