Association between cognitive status and structural brain changes in Alzheimer’s disease: Clinical implication of lightweight deep learning-aided diagnosis

Comprehensive Summary

Hsieh et al. investigate the relationship between cognitive status and structural brain changes in Alzheimer’s disease (AD) using a clinically practical deep learning model. The researchers developed a lightweight 3D convolutional neural network trained on baseline T1-weighted MRI scans from 418 AD patients and 418 age-matched cognitively normal subjects obtained from the ADNI database. The model achieved an accuracy of 90.6%, which is comparable to more complex deep learning models while using fewer parameters. Additionally, an analysis based on the occlusion map showed that the medial temporal lobe and thalamus regions contributed most to distinguishing between AD and controls, which is in accordance with AD neuropathology. Furthermore, the hierarchy regression revealed that the output from the model was able to significantly predict the variance in MMSE cognitive scores beyond demographic factors.

Outcomes and Implications

This research is significant because, currently, a biomarker is necessary to make an Alzheimer’s disease diagnosis. However, this biomarker is either very expensive or involves an invasive technique. This approach to image-based cognitive assessments is more feasible than relying on structural MRI and an efficient deep learning model. The model might also facilitate earlier diagnosis, estimation of cognitive impairment, and confidence in diagnosis in a clinical environment. While more validation in various stages of disease and involving various biomarkers is needed, the researchers suggest that, as a diagnostic tool in a clinical environment, interpretable deep learning-based models might soon be of use in clinics.