Table 2.
Comparison of our models MAE with the literature.
| Study | Model | Modality | Preprocessing | MAE | R 2 |
|---|---|---|---|---|---|
| Only using the UK Biobank | |||||
| Dartora et al. | CNN1 | T1 | Rigid reg. | 2.75 | 0.88 |
| Dartora et al. | CNN2 | T1 | Rigid reg. | 2.66 | 0.89 |
| Dartora et al. | CNN3 | T1 | Rigid reg. | 2.67 | 0.89 |
| Dartora et al. | CNN4 | T1 | Bias-field and motion cor., Skull-strip., Rigid MNI reg. | 3.03 | 0.85 |
| Bintsi et al. (2020) | CNN (3D ResNet) | T1 | Skull-strip., non-linear MNI reg. | 2.64 | 0.77 |
| Bintsi et al. (2020) | CNN (patches) | T1 | Skull-strip., non-linear MNI reg. | 2.13 | 0.85 |
| Kolbeinsson et al. (2019) | CNN | T1 | Skull-strip., non-linear MNI reg. | 2.58 | - |
| Lam et al. (2020) | R-CNN | T1 | Bias-field cor., Skull-strip., Rigid MNI reg. | 2.86 | 0.87 |
| Lam et al. (2020) | CNN | T1 | Bias-field cor., Skull-strip., Rigid MNI reg. | 4.36 | 0.66 |
| Gupta et al. (2021) | CNN (slices) | T1 | Bias-field cor., Skull-strip., Rigid MNI reg. | 2.82 | - |
| Dinsdale et al. (2021) | CNN (Female population) | T1 | UK Biobank pipeline w/linear registration to MNI | 2.86 | 0.87 |
| Dinsdale et al. (2021) | CNN (Male population) | T1 | UK Biobank pipeline w/linear registration to MNI | 3.09 | 0.86 |
| Jonsson et al. (2019) | CNN (transfer learning) | T1 | Bias-field cor., Skull-strip., Dartel MNI reg., tissue maps | 3.63 | 0.61 |
| Peng et al. (2021) | SFCN | T1 | Bias-field cor., Skull-strip., Rigid MNI reg. | 2.14 | 0.39 |
| Peng et al. (2021) | CNN | T1 | Bias-field cor., Skull-strip., Rigid MNI reg. | 2.38 | - |
| Study | Model | Modality | Samples/Cohorts | Preprocessing | MAE | R 2 |
|---|---|---|---|---|---|---|
| Using different cohorts | ||||||
| Dartora et al. | CNN1 | T1 | 16734/4 | Rigid reg. | 2.99 | 0.80 |
| Dartora et al. | CNN2 | T1 | 16734/4 | Rigid reg. | 2.67 | 0.83 |
| Dartora et al. | CNN3 | T1 | 17296/6 | Rigid reg. | 2.67 | 0.83 |
| Dartora et al. | CNN4 | T1 | 16734/4 | Bias-field and motion cor., Skull-strip., Rigid MNI reg. | 3.08 | 0.77 |
| Wood et al. (2022) | 3D DenseNet | T2 | 11735/2 | Resampling, cropping, padding, intensity normalisation. | 3.05 | - |
| Wood et al. (2022) | 3D DenseNet | T2 | 11735/2 | Resampling, cropping, padding, intensity normalisation, skull-strip. | 3.65 | - |
| Wood et al. (2022) | 3D DenseNet | T1 | 2387/1 | Skull-strip., MNI registration | 3.83 | 0.95 |
| Wood et al. (2022) | 3D DenseNet | T1 | 2387/1 | “Raw T1”* | 4.86 | 0.90 |
| He et al. (2022) | 2D CNN | T1 | 8379/8 | Bias-field cor., field of view norm., Skull-stripp., registration to SRI atlas, cropping. | 2.70 | - |
| He et al. (2021) | 3D FUS CNN | T1 split | 16705/8 | Bias-field cor., field of view norm., Skull-stripp., multi-site data harmonisation with histogram matching to SRI atlas, split in two channels (contrast and morphometric), intensity normalisation. | 3.00 | - |
| Lee et al. (2022) | 3D DenseNet | T1/PET | 4127/1 | MRI: intensity correction, and segmentation of tissues, intensity normalisation; PET: registration to MRI and MCALT space, scale to SUVRpons, segmentation in meta-ROI. | 4.20** | - |
CNN, convolutional neural network; CNN1, our CNN model in a hold-out approach; CNN2, our CNN model in a cross-validation approach; CNN3, our CNN model in a cross-validation approach, including two cohorts in the dataset; CNN4, our CNN model in a cross-validation approach, using skull-stripped images; R-CNN, recurrent CNN; SFCN, simple fully convolutional network; R2 = coefficient of determination. Dartoral et al. refer to the current study. *The authors do not specify if any processing was done. For the “raw” T2 images in the same study, resampling, cropping, and intensity normalisation were performed. **Calculated MAE only for the MRI. SUVRpons, standardised uptake volume ratio in relation to the pons; Meta-ROI, meta region of interest.