Table 1:
A list of machine learning studies about estimating brain age using healthy brain MRIs. Light gray rows are studies that included subjects spanning from youth (15 years) up to 97 years old. Dark gray rows are studies with ages spanning from pre-schoolers (3–5 years) up to 97 years old, or, in the current proposed work, from 0 to 97 years old. (Abbreviations: MAE - mean absolute error; WM - White matter; GM - Gray matter; CSF - Cerebrospinal fluid; ROI - regions of interest; CNN - Convolutional Neural network; ResNet - Residual neural network; LSTM - Long-short Term Memory;. Fus-CNN - our proposed attention-modulated multi-channel fusion CNN.)
| Study | Algorithm | #Subjects | Ages (years) | MAE (years) | Channels/feature |
|---|---|---|---|---|---|
| Part I. Studies using traditional machine learning | |||||
| Cole et al. (2017b) | Gaussian process regression | 2,001 | 18–90 | 5.01 | T1w WM/GM/CSF maps |
| Chung et al. (2018) | Ridge regressionR | 1,373 | 3–21 | 1.41 | T1w ROI features |
| Kwaket al. (2018) | Partial least square regression | 666 | 40–94 | 6.795 *% | Tiw ROI features |
| Lewis et al. (2018) | Linear regression | 1,592 | 3–22 | 1.52 | T1w GM image |
| Aycheh et al. (2018) | Gaussian process regression | 2,911 | 45–91 | 4.05 | T1w ROI features |
| Becker et al. (2018) | Gaussian process regression | 6,362 | 5–90 | 3.86 | T1w ROI features |
| Pomponio et al. (2020) | Generalized additive model | 10,477 | 3–96 | 5.35 | T1w ROI features |
| Hu et al. (2020) | Logistic regression | 251 | 0–2 | 0.09 † | T1w ROI features |
| Part II. Studies using deep learning | |||||
| Cole et al. (2017a) | 3D CNN | 2,001 | 18–90 | 4.16 | T1w image |
| Jónsson et al. (2019) | 3D ResNet | 1,264 | 15–80 | 3.63 | T1w image |
| Feng et al. (2020) | 3D CNN | 10,158 | 18–97 | 4.06 | T1w image |
| Jiang et al. (2020) | 3D CNN | 1,454 | 18–90 | 5.55 | T1w image |
| Bashyam et al. (2020) | 2D ResNet | 11,729 | 3–95 | 3.702 | T1w image |
| Peng et al. (2020) | 3D CNN | 14,503 | 42–82 | 2.14 | T1w image |
| He et al. (2020) | 2D ResNet+LSTM | 1,640 | 0–20 | 0.96 | T1w image |
| Proposed | 3D Fus-CNN | 16,705 | 0–97 | T1w split into contrast and morphometry images | |
: This study reported root mean square error (RMSE) instead of mean absolute error (MAE).
: While all other studies used cross-sectional data (MRI from only 1 visit), this study used longitudinal data, where each subject had 7 MRIs, scanned every 3 months until 1 year old and every 6 months until 2 years old.