Table 1.
Published studies about the role of radiomics in psychiatric disorders
| Study | Target | Modality | No. of cases | Brain region segmented | No. of extracted features | Feature selection technique | No. of selected features | Machine learning models | Findings |
|---|---|---|---|---|---|---|---|---|---|
| Ma et al. [19] | Differentiation between major depression disorder (MDD) and subthreshold depression (STD) | T1 weighted MRI and DTI | 142 (43 MDD, 49 STD, 50 HC), only adolescent patients | 124 regions (38 cortical and subcortical regions and 24 sulci per hemisphere) | 3,030 | RF | 25 | RF | Accuracy 86% and AUC of 0.93 for MDD vs. HC |
| Accuracy 70.5%, AUC of 0.69 for STD vs. HC | |||||||||
| Accuracy of 59%, AUC of 0.66 for MDD vs. STD | |||||||||
| Zhang et al. [20] | Differentiate between isolated Parkinson’s disease (PD) from Parkinson’s disease with depression (PDP) versus healthy volunteers (HC) | Resting state fMRI | 21 DPD, 49 PD, 50 HC | 16 gyri | 6,557 | LASSO, RF, SVM | 19 features for PDP vs. HC, 34 features for PD vs. HC, 17 features for PD vs. PDP | LASSO, RF, SVM | LASSO had highest performance. |
| Accuracy of 95 for PDP vs. HC, 96% for PD vs. HC, and 85% for PD vs. PDP | |||||||||
| Lu et al. [21] | Differentiation between Schizophrenia (SZ) vs. HC | T1 MRI, [11c]ABP-688 PET | 17 SZ, 17 HC | 5 regions: frontal cortex, posterior cingulate, temporal cortex, primary auditory cortex, and thalamus | 48 | Relief Algorithm | Between 7 to 43 for different brain regions | SVM, Bagged Trees, KNN, NB | AUC of 0.89 for T1 sequence and 0.82 for PET to differentiate SZ vs. HC |
| Xi et al. [24] | Prediction of treatment response after ECT in SZ | T1 weighted MRI | 28 responsive and 29 non-responsive patients | 19 regions | 15 first order features | t-test | 13 | Logistic regression (LR), SVM | Accuracy above 90% for LR and SVM to predict treatment response |
| Park et al. [25] | Differentiation between SZ vs. HC | T1 weighted MRI | 86 SZ, 66 HC | Bilateral hippocampi | 642 | Mutual Formation | 30 | LR, Extra-trees, AdaBoost, XGBoost, SVM | LR had highest performance: Accuracy of 82% |
| SZ patients were on treatment | |||||||||
| Cui et al. [26] | Treatment (medical treatment, ECT, and rTMS) response prediction in SZ [treatment was heterogeneous, not all cases received the same treatment] | T1 weighted MRI and resting-state fMRI | 85 responsive and 63 non-responsive patients | 91 cortical and 15 subcortical regions | 408 | LASSO | 12 | SVM | Accuracy of 80% for fMRI, 69% for T1 and 85% for the combination of T1+fMRI |
| Latha and Kavitha [27] | Differentiation between SZ vs. Schizoaffective disorder (SA) vs. HC | T1 weighted MRI | 84 HC, 81 SZ, 31 SA | Ventricle, cerebellum, and whole brain | 104 feature for ventricles, 60 features for brain, 56 for cerebellum | No feature selection | Binary Particle Swarm Optimization, Fuzzy SVM | Features of cerebellum were more accurate with accuracy of 90% | |
| Wang et al. [28] | Differentiation between bipolar disorder II (BD II) vs. HC | Resting state fMRI | 90 BD (only BD II), 117 HC | 116 regions of brain | 7,018 | Mann-Whitney U-test and LASSO | 65 | SVM | Accuracy above 80% for HC vs. BD |
| Sun et al. [32] | Differentiation of ADHD from HC subjects, and Differentiation ADHD inattentive (ADHD-I) and combined inattentive and hyperactive subtypes (ADHD-C) | T1 weighted MRI and DTI | 83 (40 with ADHD-I and 43 with ADHD-C), and 87 HC | 31 cortical regions, 7 subcortical regions per hemisphere, 24 sulci per hemisphere, 48 white matter regions | 3,106 | Forest-based | 8 feature for ADHD vs. HC and 4 features for ADHS-I vs. ADHD-C | RF | Accuracy of 73% for ADHD vs. HC, 80% for ADHD-I vs. ADHD-C |
| Kim et al. [34] | Predicting the social anxiety | Resting state fMRI | 116 | 14 regions of brain | 56 | Shapley Additive Explanation | Features of orbitofrontal cortex | Xboost, SVM, RF, and multi-layer perceptron | Xboost with highest performance; accuracy of 77% |
| Bang et al. [38] | Differentiation of panic disorder vs. HC | T1 weighted MRI | 93 patients, 120 HC | Amygdala, insula, and anterior cingulate | 1,498 | LASSO | 179 | XGBoost, RF | AUC of 0.81 for differentiation panic patients vs. HC |
| Han et al. [39] | Differentiation between Internet gaming disorder (IGD) vs. HC | 3D Fast spoiled Gradient Recalled sequence, and DTI | 59 cases, 69 HC | 101 regions of brain | 2,084 | RF | 179 | RF | Accuracy of 73% to differentiate HC vs. IGD |
| Chaddad et al. [42] | Autism spectrum disorder (ASD) vs. development control | T1 | 34 cases, 30 HC | Hippocampi and amygdala | 15 | 11 features in hippocampi and 4 features for amygdala | SVM, RF | Accuracy of 76% to differentiate autism vs. HC |
MRI, magnetic resonance imaging; DTI, diffusion tensor imaging; AUC, area under the curve; fMRI, functional MRI; LASSO, least absolute shrinkage and selection operator; RF, random forest; SVM, support vector machine; PET, positron emission tomography; KNN, k-nearest neighbors algorithm; NB, Naive Bayes algorithm; rTMS, repetitive transcranial magnetic stimulation; ADHD, attention deficit hyperactivity disorder