Table 1.
Overview of ASD studies included in this literature review.
| Authors, year, ref. | Dataset. | Sample size. | # Sites. | Input Modality. | Feature selection (y/n) | Feature engineering. | # Features. | Validation. | Best DL model. | DL ACC. | Best ML model. | ML input. | ML Acc. |
| Sen et al. 2018*, (Sen et al., 2018) | ABIDE. | ASD = 573, TD = 538. | 17. | s-MRI & rs-fMRI. | no. | Unsup. features (structural + spatio-temporal) | 45 IC for fMRI + ? sMRI. | 5 cv on training / 1 test. | multimodal feature learning + linear SVM. | 64.3. | x. | x. | x. |
| Pinaya et al. 2019**, (Pinaya et al., 2019) | HCP, ABIDE. | pretraining HC = 1113; ASD = 83, HC = 105. | 17. | s-MRI. | no. | Freesurfer cortical thickness and anatomical volumes. | x (Freesurfer 104 regions) | 10 strat cv. | AE. | 63.9a | SVM (lin) | Freesurfer cortical thickness and anatomical volumes. | 56.9 a |
| Aghdam et al. 2018, (Aghdam et al., 2018) | ABIDE I + II. | ASD = 116, TD = 69. | 7. | s-MRI & rs-fMRI. | no. | mean of AAL tc + GM/WM AAL parcellation. | 232 or 348. | 10 cv. | DBN. | 65.6. | x. | x. | x. |
| Xing et al. 2018, (Xing et al., 2018) | ABIDE I. | ASD = 527, TD = 569. | 17. | rs-fMRI. | no. | AAL (90) FC matrix. | 4005. | 10x strat 5 cv. | CNN_EW. | 66.9. | SVM. | AAL (90) FC matrix. | 63.6. |
| Ktena et al., 2018,(Ktena et al., 2018) | ABIDE. | ASD = 403, TD = 468. | 20. | rs-fMRI. | no. | anatomical spatial graphs with labels of HO FC matrix. | x. | 5 cv. | GCN. | ~67. | PCA/Euclidean. | anatomical spatial graphs with labels of HO FC matrix. | ~54. |
| Li et al. 2018, (Li et al., 2018) | ABIDE-UM. | ASD = 48, TD = 65 (+411 training) | 17* | rs-fMRI. | no. | AAL (90) FC matrix. | 4005. | strat 5 cv. | SSAE-DNN. | 67.2. | SVM. | AAL (90) FC matrix. | 60.5. |
| Kam et al. 2017, (Kam et al., 2017) | ABIDE I UM NYU. | ASD = 119, TD = 144. | 2. | rs-fMRI. | yes, hierarchical cluster! | AAL FC matrix. | x. | train/test. | DRBM. | 67.4. | SVM (graph theory) | AAL FC matrix. | 65.9. |
| Dvornek et al., 2017, (Dvornek et al., 2017) | ABIDE I. | ASD = 529, TD = 571. | 17. | rs-fMRI. | no. | CC200 tc. | 90*200. | 10 strat cv. | LSTM. | 68.5. | x. | x. | x. |
| Dvornek et al. 2018, (Dvornek et al., 2018) | 1 site. | ASD = 21, TD = 19. | 1. | task-fMRI + pheno. | no. | timeseries AAL(90) atlas. | 156*90 timeseries. | 10x 10 cv. | LSTM. | 69.8. | x. | x. | x. |
| Heinsfeld et al.2018, (Heinsfeld et al., 2018) | ABIDE I. | ASD = 505, TD = 530. | 17. | rs-fMRI. | no. | CC200 FC matrix. | 19,900. | 10 cv and leave-site out. | AE-MLP. | 70. | SVM. | CC200 FC matrix. | 65. |
| Dvornek et al. 2018, (Dvornek et al., 2018) | ABIDE I. | ASD = 529, TD = 571. | 17. | rs-fMRI + pheno. | no. | CC200 tc. | 90*200 tc + 90*5 phenotypic data. | 10 site-strat cv. | Pheno_LSTM. | 70.1. | x. | x. | x. |
| Parisot 2018, (Parisot et al., 2018) | ABIDE I. | ASD = 403, TD = 468. | 20. | rs-fMRI + pheno. | yes, RFE. | HO (1 1 0) FC matrix + pheno(sex, site) | 2000. | 10 strat cv. | GCN. | 70.4. | ridge. | HO (1 1 0) FC matrix + pheno(sex, site) | 65.3. |
| Aghdam et al. 2019, (Aghdam et al., 2019) | ABIDE I + II. | ASD = 210, TD = 249. | 20. | rs-fMRI. | no. | Max freq. voxel level. | 2D images of (~70*95) | 10 cv. | combined mixed expert CNN. | 70.5. | x. | x. | x. |
| Anirudh & Thiagarajan 2019, (Anirudh and Thiagarajan, 2017) | ABIDE I. | ASD = 403, TD = 468. | 20. | rs-fMRI. | no. | HO (1 1 0) FC matrix + pheno(sex, site) | x. | 10 cv. | ensemble G-CNN. | 70.9. | SVM(lin)* | FC matrix. | 66.8. |
| Khosla et al. 2018, (Khosla et al., 2019) | ABIDE I. | ASD=379, TD=395. | 17. | rs-fMRI. | no. | multi-channel 3D voxel connectivity maps. | x. | 10 cv (and ABIDE I/II split) | ensemble 3D CNN. | 73.5. | SVM(RBF) | FC matrix. | 71. |
| Li et al. 2018, (Li et al., 2018) | NDAR. | ASD=61, TD=215. | unclear. | s-MRI. | yes, discriminative landmarks (automatic)!! | 50 3D volumes + pheno info (sex, WB volume) | 24x24x24x50 | 10 cv | CNN | 76.24 | x | x | x |
| Mellema et al. 2019, (Mellema et al., 2019) | IMPAC | ASD=418, TD=497 | unclear | s-MRI & rs-fMRI | no | FC matrix + ROI volumes | x | Strat. 3cv | MLP | 80.4 a | Logistic Ridge Regression | FC matrix + ROI volumes | 77.34a |
| Guo et al. 2017, (Guo, 2017) | ABIDE UM | ASD=55, TD=55 | 1 | rs-fMRI | yes, based on SAE | AAL FC matrix - feature selection based on multiple SAE | 6670 | nested 5 cv | SAE-DNN | 86.4 | Elastic net | AAL FC matrix | 79,5 |
| Dekhil et al. 2018, (Dekhil, et al., 2018) | NDAR | ASD=123. TD=160 | 2 | rs-fMRI | no | PSD of tc of 34 gICA ROIs | 34*83 | 2,4,10 cv and LOO-CV with 100 permutations | SAE_SVM | 91 | PCA_SVM | PSD of tc of 34 gICA ROIs | 84 |
| Li et al. 2018, (Li et al., 2018) | 1 site | ASD=82, TD=48 | 1 | residual f-MRI | no | 2 channel (mean and std) 3D volumes | 2*32x32x32=65536 | Strat. 4 cv | 2-channel 3DCNN | 89b | RF | flattened vector of 2 channel 3D volumes (65536 dimensions) + PCA | 82b |
| Ismail et al. 2017, (Ismail, 2017) | KKI | ASD=21, TD=21 | 1 | s-MRI | yes, ROIS (automatic) | CDF of 64 shape features | 64*4000 | train/test | SAE | 92.8 | x | x | x |
| Wang et al. 2019, (Wang et al., 2019) | ABIDE I | ASD=501, TD=553 | 17 | rs-fMRI | yes, top 1000 of RFE! | AAL (1 1 6) FC matrix | 6670 | average of 5,10,20,30 cv | SVM-RFE + SSAE | 93.6 | SVM-RFE + softmax classifier | AAL (1 1 6) FC matrix | 67.3 |
* General model for ASD and ADHD, ** General model for ASD and SZ
a AUC ROC, b F score, c Balanced accuracy
! not clear if feature selection is done only on training set, !! Feature selection done before train/test split
ASD = Autism Spectrum Disorder, TD = typically developing, rs = resting state, fMRI = functional Magnetic Resonance Imaging, s-MRI = Magnetic Resonance Imaging, ABIDE = Autism Brain Imaging Data Exchange, NDAR = National Database for Autism Research, IMPAC = Maging-PsychiAtry Challenge, UM = University of Michigan, KKI = Kennedy Krieger Institute , PSD = Power Spectral Densities, Tc = timecourse, gICA = group Independent Component Analysis, NMI = Normalized Mutual Information, CDF = cumulative distribution function, WB = whole brain, PCA = principle component analysis, SVM = support vector machine, AAL = automatic anatomic labelling, CC200, craddock 200, HO = Harvard Oxford, ROI = Region of interest, CNN = convolutional neural network, EW = element-wise filter, GCN = grapch convolutional network, AE = Auto Encoder, SAE = Stacked Auto encoder, SSAE = stacked sparse auto encoder, RF = random forest, MLP = multilayer perceptron, LSTM = long short-term memory, DBN = Deep belief network, DRBM = Deep restricted Boltzmann machine, FC = functional connectivity, 10 cv = 10 fold cross validation, LOOCV = leave one out cross validation, strat cv = stratified cross validation