Alzheimer's Imaging Consortium.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Lacunes, presumed to be of vascular origin, are linked to cognitive decline and amyloid-related imaging abnormalities (ARIA).
• Accurate lacune detection is crucial for understanding neurological conditions.
• Existing methods face challenges with imbalanced datasets and distinguishing lacunes from mimicking features.

Purpose of the Study:

• To develop a deep learning model for automatic lacune segmentation.
• To enhance the distinction between lacunes and lacune-mimicking features.
• To address the challenge of imbalanced datasets in lacune detection.

Main Methods:

• Utilized 427 T2-FLAIR MRI images for model development.
• Employed an Attention U-Net architecture with an encoder pre-trained using supervised contrastive learning.
• Evaluated instance-level detection using AFROC analysis and patient-level outcomes via AUC.

Main Results:

• The model achieved a figure-of-merit (FOM) of 0.726 for instance-level lacune detection.
• Patient-level AUC for lacune detection was 0.810.
• Demonstrated moderate sensitivity in identifying patients with varying lacune counts.

Conclusions:

• A deep learning approach effectively detects lacunes in imbalanced datasets.
• Supervised contrastive learning pre-training enhances model performance.
• Future work includes regional localization and external validation.