Biomarkers.

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Imaging

Background:

• Alzheimer's disease (AD) diagnosis relies on assessing medial temporal lobe atrophy (MTA).
• Current MTA assessment uses the Scheltens scale, which is subjective and prone to inter-rater variability.
• Objective and automated MTA grading is needed for accurate early AD diagnosis and progression monitoring.

Purpose of the Study:

• To develop and validate an artificial intelligence (AI) tool for automated MTA grading.
• To eliminate inter-rater variability in MTA assessment using deep learning.
• To improve the objectivity and efficiency of MTA grading for AD diagnosis.

Main Methods:

• A dataset of 1524 MRI scans from the ADNI database and a local memory clinic was used.
• An automated MTA grading tool was developed using the YOLOv8 architecture with an EfficientViT backbone and a custom loss function.
• The AI tool's performance was compared against manual grading by experienced physicians and other deep learning models.

Main Results:

• The YOLOv8-based automated MTA grading achieved an intraclass correlation coefficient of 0.852 with manual grading (p < 0.001).
• The tool demonstrated an accuracy of 0.795, outperforming YOLOv5, Faster R-CNN, Mask R-CNN, and U-Net.
• The model showed strong generalizability across both ADNI and local memory clinic datasets.

Conclusions:

• The YOLOv8-based automated MTA grading tool accurately identifies and segments the medial temporal lobe for atrophy assessment.
• This AI approach reduces subjective variability, promoting objective and efficient MTA grading in AD diagnosis.
• The developed tool shows significant potential for clinical application in early Alzheimer's disease detection.