Improving photoacoustic imaging through the skull using deep learning: Considering 3-D effects

Abstract

Human brain mapping has become one of the most exciting fields of research for medical imaging. Among other imaging techniques, photoacoustic computed tomography (PACT) has shown great promise due to its rich optical absorption contrast, high spatial and temporal resolutions, and relatively deep penetration. However, before the acoustic signal makes it to the transducer array for image reconstruction, it becomes distorted by the porous skull, resulting in blurry images. In recent years, deep learning (DL) has been proposed as a solution to achieve images of higher quality, with U-Net being a common architecture. This numerical study utilized a large collection of blood vessel models obtained from an online database and a micro-CT scan of an exvivo human skull. Simulations were run in a 3-D space, accounting for out-of-plane acoustic phenomena. While the results show improved images, the model performs noticeably better when tested on images generated under the same conditions as the training data. These findings demonstrate U-Net’s robustness and provide insights into future architectural enhancements.