A deep ensemble hippocampal CNN model for brain age estimation applied to Alzheimer’s diagnosis

• Published in: Expert systems with applications Vol. 195; p. 116622
• Main Authors: Poloni, Katia Maria, Ferrari, Ricardo José
• Format: Journal Article
• Language: English
• Published: New York Elsevier Ltd 01.06.2022; Elsevier BV
• Subjects: Age; Age biomarker; Alzheimer's disease; Artificial neural networks; Biomarkers; Brain; Brain-age estimation; Chronology; Convolutional Neural Networks; Deep Learning; Diagnosis; Error analysis; Life expectancy; Mild cognitive impairment; Root-mean-square errors; Brain-age estimation; Convolutional Neural Networks; Age biomarker; Alzheimer’s disease; Mild cognitive impairment; Deep Learning
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2022.116622

Age-associated diseases rise as life expectancy increases. The brain presents age-related structural changes across life, with different extends between subjects and groups. During the development of neurodegenerative diseases, these changes are more intense and accentuated. As Alzheimer’s disease (AD) develops, the brain reflects accelerated aging with minor extends associated with mild cognitive impairment (MCI), i.e., the prodromal stage of AD. Therefore, it is crucial to understand a healthy brain aging process to predict a cognitive decline. This study produced an efficient age estimation framework using only the hippocampal regions that explores the associations of the brain age prediction error of age-matched cognitively normal (CN) subjects with AD and MCI subjects. For this, we have developed two convolutional neural networks. The first achieved very competitive state-of-the-art metrics, i.e., mean absolute error (MAE) of 3.31 and root mean square error (RMSE) of 4.65. The second has also achieved competitive metrics, but more importantly, we founded a statistically significant analysis of our delta estimation error between the compared groups. Further, we correlated our results with clinical measurements, e.g., Mini-Mental State Examination (MMSE) score, and obtained a significant negative correlation. In addition, we compared our results with other published studies. Therefore, our findings suggest that our delta could become a biomarker to support AD and MCI diagnosis. [Display omitted] •Hippocampal age estimation using an efficient 3D CNN architecture.•Data augmentation with an oversample over the age bins to obtain even distribution.•End-to-end framework to process new images within less than seven minutes.•Age-matched analyses with distinct aging effects and stages of neuro diseases.•Significant correlation between brain-predicted age delta error and clinical score.