Affine image registration of arterial spin labeling MRI using deep learning networks

Convolutional neural networks (CNN) have demonstrated good accuracy and speed in spatially registering high signal-to-noise ratio (SNR) structural magnetic resonance imaging (sMRI) images. However, some functional magnetic resonance imaging (fMRI) images, e.g., those acquired from arterial spin labeling (ASL) perfusion fMRI, are of intrinsically low SNR and therefore the quality of registering ASL images using CNN is not clear. In this work, we aimed to explore the feasibility of a CNN-based affine registration network (ARN) for registration of low-SNR three-dimensional ASL perfusion image time series and compare its performance with that from the state-of-the-art statistical parametric mapping (SPM) algorithm. The six affine parameters were learned from the ARN using both simulated motion and real acquisitions from ASL perfusion fMRI data and the registered images were generated by applying the transformation derived from the affine parameters. The speed and registration accuracy were compared between ARN and SPM. Several independent datasets, including meditation study (10 subjects × 2), bipolar disorder study (26 controls, 19 bipolar disorder subjects), and aging study (27 young subjects, 33 older subjects), were used to validate the generality of the trained ARN model. The ARN method achieves superior image affine registration accuracy (total translation/total rotation errors of ARN vs. SPM: 1.17 mm/1.23° vs. 6.09 mm/12.90° for simulated images and reduced MSE/L1/DSSIM/Total errors of 18.07% / 19.02% / 0.04% / 29.59% for real ASL test images) and 4.4 times (ARN vs. SPM: 0.50 s vs. 2.21 s) faster speed compared to SPM. The trained ARN can be generalized to align ASL perfusion image time series acquired with different scanners, and from different image resolutions, and from healthy or diseased populations. The results demonstrated that our ARN markedly outperforms the iteration-based SPM both for simulated motion and real acquisitions in terms of registration accuracy, speed, and generalization.