Introduction: Current deep learning-based computer-aided diagnosis (CAD) techniques face challenges in hierarchical feature extraction and computational efficiency. Traditional convolutional neural networks (CNN) often focus on local or single-scale information, neglecting global correlations of brain atrophy and multiscale pathological features. Additionally, the parameter explosion problem in deep networks limits model's generalization ability on small and medium-sized datasets. While the introduction of attention mechanisms has significantly improved feature extraction and enhanced CNN recognition capabilities, existing attention mechanisms are mostly single-scale, focusing on feature maps at specific hierarchical levels and ignoring the correlations between features of different layers.
Material and Methods: To address these issues, this study proposes a lightweight model combining a shallow feature pyramid CNN with a Dual Multi-level Attention (DMA) mechanism. Experiments using the public OASIS-1 dataset, which contains 86,437 MRI images across 4 categories, employ a focal loss function to handle class imbalance.
Results: The results show that the model including DMA outperforms both the baseline CNN and the single-scale attention mechanism in terms of accuracy (ACC), sensitivity (SEN), and specificity (SPE). Specifically, compared to CNN and CNN+CBAM: ACC improved by 3.33% and 1.26%, SEN improved by 13.2% and 0.9%, and SPE improved by 1%.
Conclusion: The model demonstrates significant advantages in distinguishing small-sample classes and differentiating between very mild dementia and normal controls, highlighting its superiority in fine-grained pathological discrimination.