Learning a Semi-Analytic Reconstruction Method for Photoacoustic Computed Tomography with Hemispherical Measurement Geometries

Photoacoustic computed tomography (PACT) is being actively developed for breast cancer imaging. In 3D PACT imagers for breast imaging, a hemispherical measurement geometry that encloses the breast has been employed. Such measurement data are referred to as “half-scan” data. Existing closed-form reconstruction methods assume a closed measurement aperture; however, the direct application of these methods to half-scan data results in inaccurate images with artifacts. Previous studies have demonstrated that half-scan data are “complete” in the sense that they contain sufficient information for accurate and stable reconstruction of an object contained within a hemispherical measurement aperture. However, direct closed-form methods for use with half-scan data have not been reported. Although optimization-based iterative image reconstruction methods are applicable, they are computationally intensive. In this work, for the first time, a semi-analytic image reconstruction method of filtered backprojection (FBP) form was proposed for use with half-scan PACT data. To accomplish this, the unknown data filtering operation is learned in a data-driven way by use of a linear U-Net neural network. To investigate the method, stochastic 3D numerical breast phantoms (NBPs) were used for model training and testing. As a result of the completeness of the half-scan data, we demonstrate that the learned FBP method can be widely applied, even when the to-be-reconstructed object differs considerably from those that were used to learn the data filtering. This is a key feature of the method that will enable it to have an important practical impact on PACT.