Deep Attention Assisted Multi-resolution Networks for the Segmentation of White Matter Hyperintensities in Postmortem MRI Scans

In the presence of cardiovascular disease and neurodegenerative disorders, the white matter of the brains of clinical study participants often present bright spots in T2-weighted Magnetic Resonance Imaging scans. The pathways contributing to the emergence of these white matter hyperintensities are still debated. By offering the possibility to directly compare MRI patterns with cellular and tissue alterations, research studies coupling postmortem imaging with histological studies are the most likely to provide a satisfactory answer to these open questions. Unfortunately, manually segmenting white matter hyperintensities in postmortem MRI scans before histology is time-consuming and laborintensive. In this work, we propose to tackle this issue with new, fully automatic segmentation tools relying on the most recent Deep Learning architectures. More specifically, we compare the ability to predict white matter hyperintensities from a registered pair of T1 and T2-weighted postmortem MRI scans of five Unet architectures: the original Unet, DoubleUNet, Attention UNet, Multiresolution UNet, and a new architecture specifically designed for the task. A detailed comparison between these five Unets and an ablation study, carried out on the sagittal slices of 13 pairs of high-resolution T1 and T2 weighted MRI scans manually annotated by neuroradiologists, demonstrate the superiority of our new approach and provide an estimation of the performance gains offered by the modules introduced in the new architecture.