Aims

Image analysis methods based on co-registration of two serial MRI scans, such as the boundary shift integral (BSI) and tensor-based morphometry (TBM), provide precise measurement of volumetric change. However, BSI is semi-automated and restricted to regions with well-defined boundaries, whereas available TBM implementations, such as Advanced Normalization Tools (ANTs), are computationally intensive and slow.

Advancements in artificial intelligence (AI) for deformable image registration allow faster computation without loss of sensitivity. Here, we present an AI-based non-linear registration algorithm to measure longitudinal brain atrophy.

Methods

Our algorithm uses a convolutional neural network and a spatial transformation network and does not require manual labelling for training. It warps one image to another by minimizing their difference, while preserving topology. Voxel-wise volume change is measured using the Jacobian Determinant of the warp field and summed within a specified region of interest. We evaluated our algorithm in ADNI1 participants (24 HC, 23 AD) and assessed the model’s ability to measure 1-year change in whole brain and ventricles. Group comparisons were performed using a Mann-Whitney U-test.

Results

Our algorithm’s registration performance was comparable to ANTs (normalized cross-correlation was 0.983 (SD=0.0065) and 0.9731 (SD=0.017) respectively). Both BSI and our algorithm showed significant group difference for whole brain and ventricles (all p < 0.003), whereas in ANTs there was a significant group effect for ventricles (p<0.001), but not whole brain (p=0.428).

Conclusions

Our algorithm produced fast and accurate non-linear registrations. Our preliminary results also show that the measured atrophy rates were sensitive to clinical diagnosis, comparable to current standard approaches.