Tau‐neurodegeneration mismatch from inter‐modality image translation using deep learning

Abstract

AbstractBackgroundThe relationship between tau neurofibrillary tangles (T) and neurodegeneration (N) may offer clues to potential presence of comorbidities, as well as relative resilience and vulnerability to Alzheimer’s disease (AD) pathology. We have previously developed measures of tau PET (T) and cortical thickness (N) mismatch based on linear model residuals. However, the underlying relationships between T and N are likely complex and non‐linear. Moreover, predicting cortical thickness based on local tau level may miss non‐local contributions of tauopathy. Here we investigate T‐N mismatch using a deep‐learning 3D image translation neural network to estimate synthetic maps of cortical thickness based on tau PET images. Deviation between the synthetic and actual cortical thickness serves as a metric of T‐N mismatch.MethodWe derived 3D cortical thickness from T1‐MRI and SUVR from 18F‐flortaucipir tracer uptake to represent N and T maps respectively. We predicted cortical thickness maps from tau SUVR by training a 3D U‐Net (Ronneberger et al. 2015) to learn the relationship between paired cortical tau SUVR and thickness maps from 70 symptomatic patients from ADNI. Approximately 70% of the training set was A+. We used regional standardized mean absolute error between predicted and actual thickness across 104 bilateral gray matter regions of interest as T‐N mismatch for clustering in a separate independent sample of 194 A+ symptomatic patients.ResultThe voxel‐wise mean absolute error of thickness translation on 194 tested patients was 0.54 mm. We obtained six T‐N data‐driven clusters (Figure 2). The group with lowest reconstruction error was defined as canonical – meaning that the degree of neurodegeneration was accurately predicted by tau PET. There were three groups with lower thickness than predicted, which were denoted as temporal‐limbic (TL) vulnerable, posterior vulnerable and diffuse vulnerable based on their spatial patterns. Additionally, two groups with greater thickness than predicted were defined as anterior resilient and diffuse resilient. These groups differed in subsequent longitudinal cognitive decline measured by CDR‐SB (Figure 3). The vulnerable groups displayed greater cognitive decline than the canonical group, which may be attributable to associations with comorbidities.ConclusionOur findings support a deep‐learning T‐N mismatch approach for disentangling AD heterogeneity.