Background Although convolutional neural networks (CNNs) achieve high diagnostic accuracy for detecting Alzheimer’s disease (AD) dementia based on magnetic resonance imaging (MRI) scans, they are not yet applied in clinical routine. One important reason for this is a lack of model comprehensibility. Recently developed visualization methods for deriving CNN relevance maps may help to fill this gap as they allow the visualization of key input image features that drive the decision of the model. We investigated whether models with higher accuracy also rely more on discriminative brain regions predefined by prior knowledge. Methods We trained a CNN for the detection of AD in N = 663 T1-weighted MRI scans of patients with dementia and amnestic mild cognitive impairment (MCI) and verified the accuracy of the models via cross-validation and in three independent samples including in total N = 1655 cases. We evaluated the association of relevance scores and hippocampus volume to validate the clinical utility of this approach. To improve model comprehensibility, we implemented an interactive visualization of 3D CNN relevance maps, thereby allowing intuitive model inspection. Results Across the three independent datasets, group separation showed high accuracy for AD dementia versus controls (AUC ≥ 0.91) and moderate accuracy for amnestic MCI versus controls (AUC ≈ 0.74). Relevance maps indicated that hippocampal atrophy was considered the most informative factor for AD detection, with additional contributions from atrophy in other cortical and subcortical regions. Relevance scores within the hippocampus were highly correlated with hippocampal volumes (Pearson’s r ≈ −0.86, p < 0.001). Conclusion The relevance maps highlighted atrophy in regions that we had hypothesized a priori. This strengthens the comprehensibility of the CNN models, which were trained in a purely data-driven manner based on the scans and diagnosis labels. The high hippocampus relevance scores as well as the high performance achieved in independent samples support the validity of the CNN models in the detection of AD-related MRI abnormalities. The presented data-driven and hypothesis-free CNN modeling approach might provide a useful tool to automatically derive discriminative features for complex diagnostic tasks where clear clinical criteria are still missing, for instance for the differential diagnosis between various types of dementia.

Objective:We assessed whether novelty-related fMRI activity in medial temporal lobe (MTL) regions and precuneus follows an inverted U-shape pattern across the clinical spectrum of increased Alzheimer disease (AD) risk as previously suggested. Specifically, we tested for potentially increased activity in individuals with higher AD risk due to subjective cognitive decline (SCD) or mild cognitive impairment (MCI). We further tested whether activity differences related to diagnostic groups were accounted for by CSF markers of AD or brain atrophy.Methods:We studied 499 participants aged 60-88 years from the German Center for Neurodegenerative Diseases Longitudinal Cognitive Impairment and Dementia Study (DELCODE) who underwent task-fMRI. Participants included 163 cognitively normal (healthy control, HC) individuals, 222 SCD, 82 MCI, and 32 patients with clinical diagnosis of mild AD. CSF levels of β-amyloid 42/40 ratio and phosphorylated-tau181 were available from 232 participants. We used region-based analyses to assess novelty-related activity (novel > highly familiar scenes) in entorhinal cortex, hippocampus and precuneus, as well as whole-brain voxel-wise analyses. First, general linear models tested differences in fMRI activity between participant groups. Complementary regression models tested quadratic relationships between memory impairment and activity. Second, relationships of activity with AD CSF biomarkers and brain volume were analyzed. Analyses were controlled for age, gender, study site, and education.Results:In the precuneus, we observed an inverted U-shape pattern of novelty-related activity across groups, with higher activity in SCD and MCI compared to HC, but not in AD patients who showed relatively lower activity than MCI. This non-linear pattern was confirmed by a quadratic relationship between memory impairment and precuneus activity. Precuneus activity was not related to AD biomarkers or brain volume. In contrast to precuneus, hippocampal activity was reduced in AD dementia compared to all other groups and related to AD biomarkers.Conclusion:Novelty-related activity in the precuneus follows a non-linear pattern across the clinical spectrum of increased AD risk. While the underlying mechanism remains unclear, increased precuneus activity might represent an early signature of memory impairment. Our results highlight the non-linearity of activity alterations that should be considered in clinical trials using functional outcome measures or targeting hyperactivity.

We investigated whether the impact of tau-pathology on memory performance and on hippocampal/medial temporal memory function in non-demented individuals depends on the presence of amyloid pathology, irrespective of diagnostic clinical stage. We conducted a cross-sectional analysis of the observational, multicentric DZNE-Longitudinal Cognitive Impairment and Dementia Study (DELCODE). Two hundred and thirty-five participants completed task functional MRI and provided CSF (92 cognitively unimpaired, 100 experiencing subjective cognitive decline and 43 with mild cognitive impairment). Presence (A+) and absence (A−) of amyloid pathology was defined by CSF amyloid-β42 (Aβ42) levels. Free recall performance in the Free and Cued Selective Reminding Test, scene recognition memory accuracy and hippocampal/medial temporal functional MRI novelty responses to scene images were related to CSF total-tau and phospho-tau levels separately for A+ and A− individuals. We found that total-tau and phospho-tau levels were negatively associated with memory performance in both tasks and with novelty responses in the hippocampus and amygdala, in interaction with Aβ42 levels. Subgroup analyses showed that these relationships were only present in A+ and remained stable when very high levels of tau (&gt;700 pg/ml) and phospho-tau (&gt;100 pg/ml) were excluded. These relationships were significant with diagnosis, age, education, sex, assessment site and Aβ42 levels as covariates. They also remained significant after propensity score based matching of phospho-tau levels across A+ and A− groups. After classifying this matched sample for phospho-tau pathology (T−/T+), individuals with A+/T+ were significantly more memory-impaired than A−/T+ despite the fact that both groups had the same amount of phospho-tau pathology. ApoE status (presence of the E4 allele), a known genetic risk factor for Alzheimer’s disease, did not mediate the relationship between tau pathology and hippocampal function and memory performance. Thus, our data show that the presence of amyloid pathology is associated with a linear relationship between tau pathology, hippocampal dysfunction and memory impairment, although the actual severity of amyloid pathology is uncorrelated. Our data therefore indicate that the presence of amyloid pathology provides a permissive state for tau-related hippocampal dysfunction and hippocampus-dependent recognition and recall impairment. This raises the possibility that in the predementia stage of Alzheimer’s disease, removing the negative impact of amyloid pathology could improve memory and hippocampal function even if the amount of tau-pathology in CSF is not changed, whereas reducing increased CSF tau-pathology in amyloid-negative individuals may not proportionally improve memory function.