Assessment of Convolutional Neural Networks for Dementia Diagnosis through MR Images

Dementia is marked by a significant decline in cognitive abilities, surpassing typical age-related changes. This decline spans various cognitive domains, including memory, thinking, orientation, comprehension, calculation, learning, language, and judgment. In this study, we evaluated 15 Convolutional Neural Network (CNN) models that were trained to detect dementia using the OASIS (Open Access Series of Imaging Studies) dataset, which consists of MR images classified into four dementia categories. The densenet201 model outperformed the others, achieving an accuracy of 86.53%, with a sensitivity and specificity of 86.53% and 95.51%, respectively. Precision was recorded at 87.75%, with a false positive rate (FPR) of 4.49%, and an F1 Score of 86.27%. Furthermore, the Matthews Correlation Coefficient (MCC) was 82.54%, and the Kappa coefficient was 64.07%. Consequently, densenet201 stood out as the most effective model for dementia diagnosis from MR images, showcasing robust performance across various metrics.