Multi-Architecture Deep Learning for Early Alzheimer's Detection in MRI: Slice- and Scan-Level Analysis

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Alzheimer's disease (AD) is a progressive neurodegenerative disorder, with early detection crucial for intervention.
• Structural MRI reveals brain changes in AD, but manual detection of subtle, early-stage alterations is challenging.
• Automated methods are needed for reliable diagnosis, especially during prodromal stages like mild cognitive impairment (MCI).

Purpose of the Study:

• To compare fifteen deep learning architectures for classifying AD, MCI, and cognitively normal (CN) subjects.
• To investigate the impact of preprocessing pipelines on model reliability and performance.
• To establish preprocessing quality as a key determinant for deep learning in neuroimaging.

Main Methods:

• A dual-level evaluation framework (slice-level and scan-level) was used.
• Fifteen deep learning models, including CNNs, Transformers, and hybrid architectures, were tested on the ADNI dataset.
• Preprocessing steps were systematically optimized to reduce data variability and enhance feature consistency.

Main Results:

• CNNs and hybrid pre-trained models outperformed Transformer-based models in both slice-level and scan-level classification.
• ConvNeXtV2-L achieved the highest scan-level performance (91.07%), and EfficientNetV2-L the highest slice-level accuracy (86.84%).
• Optimized preprocessing significantly influenced model performance, with some architectures showing substantial improvements and others degradation, indicating interdependence between architecture and aggregation strategy.

Conclusions:

• Robust and standardized preprocessing is essential for reliable deep learning in neuroimaging.
• Carefully designed preprocessing enhances classification accuracy for AD detection.
• This work provides a foundation for more reproducible and interpretable Alzheimer's disease detection pipelines.