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

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Radiology

Background:

• Voxel-based morphometry (VBM) detects brain atrophy in structural MRI for neurodegenerative disease diagnosis.
• Conventional VBM is sensitive to MRI scanner variations and acquisition parameters.
• A novel convolutional neural network-based VBM (CNN-VBM) was developed, independent of normative reference databases.

Purpose of the Study:

• To clinically validate the performance of CNN-based VBM.
• To compare CNN-VBM with conventional VBM in patients with suspected neurodegenerative disease.

Main Methods:

• CNN-VBM was evaluated against conventional VBM in 227 patients with suspected neurodegenerative disease.
• VBM maps were visually assessed by two readers for disease detection and differentiation of atrophy patterns (e.g., Alzheimer's disease).
• Simultaneously acquired 18F-fluorodeoxyglucose Positron Emission Tomography (FDG-PET) served as the reference standard.

Main Results:

• CNN-VBM significantly impacted the visual detection of neurodegenerative disease (p < 0.001).
• CNN-VBM achieved higher balanced accuracy (80.4% vs. 75.7%), sensitivity (86.3% vs. 79.5%), and specificity (74.5% vs. 71.8%) compared to conventional VBM.
• No significant difference was found in differentiating Alzheimer's disease-typical atrophy patterns between the two VBM methods (p = 0.871).

Conclusions:

• CNN-based VBM offers clinically relevant accuracy for detecting neurodegeneration-suspect atrophy.
• CNN-VBM demonstrates superior sensitivity compared to conventional VBM using a mixed-scanner normative database.
• The CNN-based approach maintains specificity, providing a robust alternative for neuroimaging analysis.