Utilizing Variational Autoencoders in the Bayesian Inverse Problem of Photoacoustic Tomography

Utilizing variational autoencoders in photoacoustic tomography

“…In our previous work, we trained the neural network Λ to estimate the posterior distribution based on a set of photoacoustis measurement data p t . 15 In this work, we extend our previous work by estimating the posterior distribution based on a time reversal reconstruction of the initial pressure…”

“…20 A detailed description of the vessel phantoms and data simulation can be found from Ref. 15 In short, the simulated photoacoustic data p were computed using the noisy data samples p (i) t . Altogether, 50000 quadruples of simulated data, TR-reconstructions, true initial pressure images, and noise levels were used.…”

“…Recently, we proposed utilizing the UQ-VAE in solving the Bayesian inverse problem of PAT based on a set of measured photoacoustic data. 15 In this work, we extend the work by formulating the UQ-VAE as a post-processing approach where the posterior distribution is estimated from a time-reversal (TR) reconstruction. The neural network is trained by minimizing a loss function derived from minimization of a divergence between the true and approximate posterior distributions.…”

“…The neural network is trained by minimizing a loss function derived from minimization of a divergence between the true and approximate posterior distributions. [14][15][16] Furthermore, both data consistency term and physical ultrasound model are utilized in the training process.…”

Deep learning in photoacoustic tomography utilizing variational autoencoders

“…In this work, we propose a deep learning approach for the Bayesian inverse problem of PAT based on the uncertainty quantification variational autoencoder (UQ-VAE). 5 The approach is evaluated using 2D simulations in a limited view sensor geometry and the results are compared to the solution of the inverse problem in a Bayesian framework.…”