Deep-learning framework and acoustic reflector for improved limited-view and sparse photoacoustic tomography

Abstract

Photoacoustic imaging is a hybrid imaging modality that relies upon optical absorption of pulsed light and subsequent thermoelastic generation of ultrasound. The detection of the induced acoustic waves outside the tissue enables image reconstruction. A major challenge encountered in photoacoustic tomography (PAT) lies in the inability to acquire complete projection data from the region of interest due to both the limited view and sparsity of available ultrasonic sensors. The resulting images are characterized by severe artifacts and poor quality. In this work, we examined the utility of incorporating an acoustic reflector to address the limited view problem and to train a convolutional neural network (CNN) to improve PAT image reconstruction from sparsely sampled data. Photoacoustic wave propagation was simulated in MATLAB using the k-Wave toolbox. We compared the performance of a sparse linear transducer array (with and without reflector) to that of a circular transducer array. The structural similarity index (SSI) was used as a metric for evaluating image quality. The combination of a curved reflector and artifact-removal using a CNN improved the quality of PAT images from the linear configuration. The resulting mean SSI value (0.859) was comparable to that achieved using the circular transducer array (0.926).