Using denoising diffusion probabilistic models to enhance quality of limited-view photoacoustic tomography

In photoacoustic tomography (PAT), a limited angle of detector coverage around the object affects PAT image quality. Consequently, PAT images can become challenging to interpret accurately. Although deep learning methods, such as convolutional neural networks (CNNs), have shown promising results in recovering high-quality images from limited-view data, these methods suffer from loss of fine image details. Recently, denoising diffusion probabilistic models (DDPM) are gaining interest in image generation applications. Here we explore the potential of conditional denoising diffusion probabilistic models (cDDPM) to enhance quality of limited-view PAT images. The OADAT dataset consisting of 2D PAT images of healthy forearms acquired with a semicircle array of 256 ultrasound elements is used. PAT images are reconstructed using the full array (256 elements) and also the central 128, 64 and 32 elements for limited-view. The approach showed to be able to filter out limited-view streak artifacts, accurately recover shapes of vascular structures, and preserve fine-detailed texture. Conditional DDPMs show potential in improving quality of limited-view PAT reconstructions, however, they require higher computational cost compared to conventional CNNs. Future works will include the reduction of computational time and further evaluations on different datasets and array geometries.