Self-Guided and MR-Guided Deep-Learned Post-Reconstruction PET Processing

Reconstructed PET images exhibit high noise levels and low spatial resolution when shorter scan times and reduced injected doses are used. Regularisation methods such as post-reconstruction smoothing can help to improve image quality. Recently, neural networks have proved to be highly effective for this task by learning an ensemble of kernels. For post-processing of PET images, a high-resolution MR image can also be used for guidance to further improve the final image quality. In this work, we investigate the impact of the input choice and the number of training samples used on a neural network's performance for PET post-reconstruction processing. To do so, six combinations of low-count PET and MR independent reconstruction outputs are fed into a state-of-the-art residual convolutional neural network (CNN). Six different networks were trained using as input i) the last iteration of a conventional PET reconstruction, ii) all the iterates from the PET reconstruction, iii) only the final PET and MR estimates, iv) all the PET estimates and the final MR, v) the final PET and all the MR estimates and vi) all the iteration outputs of independent PET and MR reconstruction. The networks have been trained using a different number of training samples as well. The results obtained suggest that using all the intermediate reconstructions lead the network to perform better when the training set size is limited. Furthermore, the gain in performance observed when the dataset size increases are higher for methods using all the intermediate reconstruction outputs. Future work will focus on training networks with a higher number of training samples to confirm the trend observed and assess the proposed method on 3D real data.