A comparative study on 2-D convolutional neural network architectures for the detection of cognitive impairment from structural MRI

Cognitive impairment (CI), ranging from very mild to moderate severity, represents a clinically meaningful indicator of neurodegenerative decline and and frequently precedes the onset of Alzheimer’s disease (AD). Structural MRI serves as a non-invasive biomarker capable of capturing subtle atrophic changes associated with different stages of cognitive impairment. In this study, we present a comparative analysis of several 2D convolutional neural network (CNN) architectures for detecting different stages of cognitive impairment from T1-weighted structural MRI slices. Among the evaluated architectures, ResNet-50 achieved the best overall performance with an accuracy of 0.97 and macro-average F1-score of 0.97, demonstrating strong feature extraction capabilities for stages of cognitive impairment. The ensemble of models has been examined and proved to serve as an additional validation step demonstrating that multiple architectures converge toward similar predictions. Grad-CAM visualizations have established that the models consistently attended clinically relevant regions, including hippocampal atrophy, enlargement of the inferior and lateral ventricles, and thinning around the parietal-temporal cortices. The results indicate that optimized 2D CNN frameworks can effectively differentiate stages of cognitive impairment from structural MRI, a task that is clinically meaningful because CI staging reflects early neurodegenerative changes commonly associated with Alzheimer’s disease, while also offering the practical advantage of reduced computational complexity compared to 3D CNN models.