Learning-Based Attenuation Correction for Brain PET/MRI Using Artificial Neural Networks

The purpose of this study is to build a learning-based attenuation map estimation scheme for brain attenuation correction of the positron emission tomography/magnetic resonance (PET/MR) imaging using an artificial neural network (ANN). The attenuation map estimation is cast as a regression problem that models a nonlinear mapping between the MR image patches and the corresponding patches of the attenuation map. An ANN is used to solve the regression problem through learning from examples. Using the BrainWeb phantoms, we simulated brain PET data with corresponding MR image and attenuation map pairs of 20 subjects. The ANN was trained with the image patches from the MR image and attenuation map of 1 subject. The trained ANN was then applied to estimate the attenuation maps from the MR images for the other 19 subjects. The estimated attenuation maps were compared with their simulated counterparts using the mean absolute error (MAE) on regions of soft tissue and bone. The impact of estimation accuracy on the reconstructed PET images was evaluated by calculating the relative error with respect to the PET images reconstructed using the AC with the simulated attenuation maps. The estimated attenuation maps obtain the MAE of 0.0021 for soft tissue and 0.0096 for bone, revealing strong agreement with the corresponding simulated ones. The images reconstructed with the estimated attenuation maps have average relative errors of -0.24%± 0.8% on gray matter and of -0.28%± 0.28% on white matter. We demonstrate that the ANN model trained using the MR image and attenuation map of one subject applies well to other subjects and the estimated attenuation map has potential to produce accurate AC for brain PET/MR imaging.