Clinical validation of artificial intelligence-based single-subject morphometry without normative reference database.

Background: Single-subject voxel-based morphometry (VBM) is a powerful technique for reader-independent detection of brain atrophy in structural magnetic resonance imaging (MRI) to support the (differential) diagnosis and staging of neurodegenerative diseases in individual patients. However, VBM is sensitive to the MRI scanner platform and details of the acquisition sequence. To mitigate this limitation, we recently proposed and validated a convolutional neural network (CNN)-based VBM which does not rely on a normative reference database.

Objective: Clinical validation of CNN-based VBM.

Methods: CNN-based VBM was compared with conventional VBM based on a mixed-scanner normative database in 227 consecutive patients (66.0 ± 9.6 years, 53.3% female) with suspected dementing neurodegenerative disease. VBM maps were interpreted visually by two experienced readers, first with respect to the presence of any neurodegenerative disease, then for the differentiation between Alzheimer's disease (AD)-typical and non-AD atrophy patterns. A Likert 6-score was used for both tasks. Simultaneously acquired positron emission tomography (PET) with F-fluorodeoxyglucose (FDG) served as reference standard.

Results: Repeated-measures ANOVA revealed a significant impact of the VBM method on the visual detection of any neurodegenerative disease (p < 0.001). Balanced accuracy/sensitivity/specificity were 80.4/86.3/74.5% for CNN-based VBM versus 75.7/79.5/71.8% for conventional VBM. Differentiation between AD and non-AD typical atrophy patterns did not differ between both VBM methods (p = 0.871).

Conclusions: CNN-based VBM provides clinically useful accuracy for the detection of neurodegeneration-suspect atrophy with higher sensitivity than conventional VBM with a mixed-scanner normative reference database and without compromising specificity.