Enhancing Bayesian PET image reconstruction using neural networks

The performance of smoothness-enforced Bayesian PET image reconstruction is strongly affected by the weight on regularization. Compromises need to be made between variance and spatial resolution. In this work, we propose to use an artificial neural network (ANN) to fuse the image versions reconstructed from a maximum a posteriori (MAP) algorithm with different regularizing weights for quantitative improvement. Using the BrainWeb phantoms, we simulated PET data at different count levels for different subjects with and without lesions. We designed an ANN and trained it using MAP reconstructions with different regularization parameters for one normal subject at a specific count level. Reconstructed images from the simulations with lesions, of other count levels, and of other subjects were fused using the trained ANN. In all of the testing experiments, the designed ANN fusion keeps the noise level as low as what the MAP algorithm achieves at heavy regularization while significantly reduces the bias or improves the lesion contrast. We conclude that the proposed ANN fusion method removes the need for tuning the regularization of the MAP algorithm.