Deep networks for behavioral variant frontotemporal dementia identification from multiple acquisition sources.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Neurology

Background:

• Diagnosing behavioral variant frontotemporal dementia (bvFTD) is challenging, particularly in early stages.
• Current diagnostic criteria rely on MRI showing frontal and anterior temporal lobe atrophy, but processing is device-dependent and requires manual feature extraction.

Purpose of the Study:

• To develop and evaluate artificial neural networks for accurate bvFTD identification using MRI data.
• To assess the generalization and stability of deep learning models across different MRI acquisition devices without fine-tuning.

Main Methods:

• Utilized various classes of artificial neural networks, including attention-based deep networks.
• Performed extensive hyperparameter searches to optimize model performance.
• Employed data intra-mixing to improve model robustness across different acquisition devices.

Main Results:

• Achieved classification accuracy exceeding 90% on both Area Under the Receiver Operating Characteristic Curve (AuROC) and balanced accuracy metrics.
• Demonstrated model stability and generalization, identifying bvFTD even with inter-device MRI data.
• Showcased the ability of models to perform without device-specific fine-tuning.

Conclusions:

• Deep learning, particularly attention-based networks, offers a robust and accurate method for bvFTD diagnosis from MRI.
• This approach overcomes device-dependency issues in neuroimaging analysis for bvFTD.
• The developed models show significant potential for improving early and reliable diagnosis of bvFTD.