Development and Validation of a Deep Learning-Based Automatic Brain Segmentation and Classification Algorithm for...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Deep learning shows potential for predicting Alzheimer disease from T1-weighted brain MR images.
• Existing methods for Alzheimer disease diagnosis can be limited.

Purpose of the Study:

• To develop and validate a deep learning algorithm for Alzheimer disease diagnosis.
• Utilize 3D T1-weighted brain MR images for automated segmentation and classification.

Main Methods:

• Developed a 2-step deep learning algorithm using convolutional neural networks for brain parcellation.
• Employed XGBoost classifier, compared with logistic regression and Support Vector Machine, for disease prediction.
• Utilized 5-fold cross-validation across four datasets totaling 2637 patients.

Main Results:

• XGBoost significantly improved Alzheimer disease prediction compared to other methods (P < .001).
• For differentiating Alzheimer disease from mild cognitive impairment, XGBoost achieved an AUC of 0.825 with 68% sensitivity and 70% specificity.
• For differentiating mild cognitive impairment from healthy controls, XGBoost achieved an AUC of 0.870 with 79% sensitivity and 80% specificity.

Conclusions:

• The deep learning algorithm enables accurate Alzheimer disease diagnosis from T1-weighted brain MR images.
• The algorithm's reliance on widely available T1-weighted MRI makes it a promising tool for widespread Alzheimer disease prediction.