Voxel-based morphometry in single subjects without a scanner-specific normal database using a convolutional neural...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Voxel-based morphometry (VBM) is crucial for detecting disease-specific brain atrophy in MRI scans.
• Conventional VBM requires scanner-specific normal databases (NDB), limiting its widespread clinical application.
• The absence of readily available NDBs presents a significant barrier to the routine use of VBM.

Purpose of the Study:

• To design, train, and validate a deep convolutional neural network (CNN) for single-subject VBM analysis.
• To develop a CNN-based VBM (CNN-VBM) method that eliminates the need for scanner-specific NDBs.
• To assess the performance of CNN-VBM in detecting atrophy associated with Alzheimer's disease (AD) and frontotemporal lobar degeneration (FTLD).

Main Methods:

• A CNN was trained on 8945 T1w MRI scans from 65 scanners, using conventional VBM with scanner-specific NDBs for gold standard maps.
• CNN-VBM was evaluated on an independent dataset of healthy controls, AD patients, and FTLD patients.
• Performance was quantified using Dice coefficients and visual categorization accuracy by two independent readers.

Main Results:

• CNN-VBM maps demonstrated high similarity to conventional VBM maps, with a median Dice coefficient of 0.85.
• The overall accuracy for visually categorizing VBM maps to detect AD or FTLD was 89.8% for CNN-VBM and 89.0% for conventional VBM.
• The performance of CNN-VBM was comparable to traditional VBM methods.

Conclusions:

• CNN-VBM effectively detects disease-specific atrophy without the necessity of scanner-specific normal databases.
• This deep learning approach offers a viable alternative to conventional VBM, overcoming the limitation of NDB availability.
• CNN-VBM has the potential to facilitate the broader clinical adoption of VBM for diagnosing neurodegenerative conditions.