Concurrent 3D super resolution on intensity and segmentation maps improves detection of structural effects in neurodegenerative disease

We propose a new perceptual super resolution (PSR) method for 3D neuroimaging and evaluate its performance in detecting brain changes due to neurodegenerative disease. The method, concurrent super resolution and segmentation (CSRS), is trained on volumetric brain data to consistently upsample both an image intensity channel and associated segmentation labels. The simultaneous nature of the method improves not only the resolution of the images but also the resolution of associated segmentations thereby making the approach directly applicable to existing labeled datasets. One challenge to real world evaluation of SR methods such as CSRS is the lack of high resolution ground truth in the target application data: clinical neuroimages. We therefore evaluate CSRS effectiveness in an adjacent, clinically relevant signal detection problem: quantifying cross-sectional and longitudinal change across a set of phenotypically heterogeneous but related disorders that exhibit known and differentiable patterns of brain atrophy. We contrast several 3D PSR loss functions in this paradigm and show that CSRS consistently increases the ability to detect regional atrophy both longitudinally and cross-sectionally in each of five related diseases.