Experimental Validation of Learning-based Compensation for Skull-induced Aberrations in Transcranial Photoacoustic Computed Tomography

Transcranial photoacoustic computed tomography (PACT) holds great potential as a neuroimaging modality, but there remains a need to develop practical and effective image reconstruction methods. Model-based reconstruction methods can compensate for these skull-induced aberrations but are computationally intensive, often requiring tens of hours or even days to reconstruct a single image using contemporary hardware. Additionally, these methods rely on precise knowledge of the skull’s elastic and acoustic parameters, which is generally unavailable. This study investigates a learning-based image reconstruction method for 3D transcranial PACT that is robust to modeling errors arising from uncertainty in the skull’s acoustic and elastic properties. The method also has the potential to reduce image reconstruction time by two orders of magnitude compared to traditional model-based approaches. In the reconstruction process, a preliminary image is first computed using the adjoint of the imaging forward operator, a computationally efficient yet approximate method. A U-Net-based deep neural network is then employed to map the preliminary image into a high-quality, de-aberrated estimate of the induced initial pressure distribution within the cortical region of the brain. The proposed method was experimentally validated using a physical phantom containing an adult human skull. Results show that the learning-based method achieved image quality comparable to a finely tuned optimization-based method while reducing computational time from 30 hours to 10 minutes. This is the first experimental demonstration of a learned image reconstruction method for 3D transcranial PACT, marking a significant advancement in the field.