Association between cognitive status and structural brain changes in Alzheimer's disease: Clinical implication of...

The complex brain changes involved in Alzheimer's disease (AD) development constitute a high-dimensional nonlinear feature space where deep learning (DL) classification/diagnosis may be advantageous over classical non-learning methods. However, the practicality of DL remains under debate among healthcare professionals, largely because many models are computationally expensive and operate without explicit interpretability. This study aimed to construct a lightweight DL model to disclose the association between cognitive status and structural brain changes in AD.

By using the data obtained from the Alzheimer's Disease Neuroimaging Initiative database, 418 AD patients and 418 age-matched cognitively normal (CN) subjects were included for DL model construction based on their T1-weighted magnetic resonance images at baseline visit. A lightweight design was achieved by incorporating group convolution, global pooling, and efficient channel attention.

The accuracy rate of our model was 90.6 %, competitive with previous models built with up-to-ten times more parameters. The occlusion maps showed that the medial temporal area and thalamus accounted the most for our model's differentiation between AD and CN, in line with current knowledge of the pathological trajectory. Hierarchical regression further revealed that the logit of the DL model output explained a significant amount of variance in the mini mental state examination score, above and beyond the clinical indices including age, sex, and education duration (R² change = 0.341, F(1, 91) = 57.623, p < 0.001).

Lightweight DL can be clinically practicable for AD diagnosis by focusing on pathologically interpretable structural changes and offering image-based assessment of cognitive status.