Model-Based Speed-of-Sound Reconstruction via Interpretable Pruned Priors

Image reconstruction problems for estimating parametric maps do not typically allow for large annotated in vivo datasets, required by traditional black-box deep learning models. Model-based deep learning allows the integration of known imaging models, where variational network offers interpretable priors that can be learned merely from simulations while still generalizable to actual data despite large domain gaps. Such learned priors, however, are often overparameterized, resulting in large networks, with potential for overfitting and difficulty in the analysis of priors. In this paper, we introduce pruning in modelbased deep learning with variational networks for distilling the learned priors, using two pruning strategies based on magnitudes and correlations on a sample set. Priors are pruned in multiple steps until a desired sparsity is achieved. We demonstrate this in computed tomography, for the image reconstruction of speed-of-sound maps with ultrasound. Our quantitative evaluation shows comparable performance of networks with 32 priors and pruned 2 priors in simulation and phantom experiments. Our results also demonstrate the generalizability of networks trained only on simple simulated data to real data acquisitions.