Source-linked AI summary
Structural neuroimaging as clinical predictor: a review of machine learning applications
José María Mateos-Pérez, Mahsa Dadar, María Lacalle-Aurioles, Yasser Iturria-Medina, Yashar Zeighami, Alan C. Evans
TL;DR
Structural MRI machine-learning studies need a synthesis of their predictive methods, practical problems, and disease applications. This review surveys those areas across disorders, finding high reported accuracy for distinguishing controls from patients but lower accuracy for progressive-versus-stable classifications, while noting the need for double-blind clinical prediction experiments.
Problem
Group-level imaging differences do not necessarily provide accurate individual clinical predictions, motivating a review focused on predictive classification rather than correlation.
Method
The paper reviews machine-learning techniques applied to structural MRI, covering practical issues, predictive evaluation, and applications across diseases.
Results
Reported accuracies are higher than 0.9 for differentiating controls from patients and generally around 0.7 for distinguishing progressive from stable patients.
Takeaways & Limitations
The review provides benchmarks for structural-MRI classification and highlights the value of examining practical application across diseases and disorders.
Takeaways & Limitations
The reviewed studies provide benchmarks, but practical clinical use still requires double-blind experiments with predictions made before clinical outcomes.
Abstract
from arXiv · showhide
In this paper, we provide an extensive overview of machine learning techniques applied to structural magnetic resonance imaging (MRI) data to obtain clinical classifiers. We specifically address practical problems commonly encountered in the literature, with the aim of helping researchers improve the application of these techniques in future works. Additionally, we survey how these algorithms are applied to a wide range of diseases and disorders (e.g. Alzheimer's disease (AD), Parkinson's disease (PD), autism, multiple sclerosis, traumatic brain injury, etc.) in order to provide a comprehensive view of the state of the art in different fields.
1. Introduction
This review examines machine-learning classifiers built from structural MRI, emphasizing out-of-sample clinical prediction rather than group-level correlations. It surveys applications across diseases while addressing practical and interpretability constraints.
- Scope and inputs: Structural MRI studies use measures such as cortical thickness, gray-matter density, white-matter microstructure, connectivity, and network metrics as machine-learning inputs.Features may be computed per voxel or summarized across anatomical regions to reduce dimensionality.
- Scope and aims: The review covers classifiers that predict clinical states and identify brain regions related to disease.It focuses on structural MRI, while referring to other modalities only when generalizations apply.
- Motivation: Group-level differences do not necessarily yield accurate predictions for individual subjects, so imaging metrics should be assessed on new data samples.The review therefore excludes papers providing only correlational analyses.
- Evaluation focus: The review defines prediction through out-of-sample testing, including cross-validation, because reported accuracy is more likely to generalize beyond training data.The reviewed methods focus on relatively small feature sets compared with the number of image voxels.
- Scope boundaries: Deep-learning approaches are excluded because they generally require larger datasets and more computational power and present interpretability challenges.The paper describes these techniques as typically regarded as black boxes.
- Scope boundaries: The introduction does not provide detailed explanations of individual machine-learning algorithms and instead directs readers to existing reviews and textbooks.Referenced methods include support vector machines, linear discriminants, logistic regression, and random forests.
2. From imaging to prediction: an overview
The paper outlines a pipeline that converts processed imaging data into subject-level feature matrices and clinical labels, then trains and evaluates predictive models. It also discusses cross-validation, ensembling, and the trade-off between accuracy and interpretability.
- Image processing: Structural and diffusion MRI are processed into quantitative measures such as cortical thickness, gray-matter density, fractional anisotropy, connectivity matrices, and network metrics.Preprocessing can include denoising, intensity correction, registration, segmentation, and cortical-surface extraction.
- Feature extraction: Feature extraction removes spatial information by flattening computed matrices into one numerical feature vector per subject.For 40,000 cortical vertices, cortical-thickness values produce a 40000×1 vector.
- Feature extraction: Subject-level feature vectors are aggregated into an N × M matrix, optionally including demographic or behavioral information, with clinical states forming the target labels.N denotes the number of subjects and M the feature-vector length.
- Predictive modeling: A predictive classifier learns the relationship between input vectors and categorical clinical labels, but must be tested on data not used during training.Imaging datasets are often too scarce to reserve a separate test set.
- Predictive modeling: Cross-validation partitions the dataset into folds, trains on N − 1 folds, tests on the remaining fold, and repeats until every fold has served as the test set.Reported metrics can include accuracy, sensitivity, and specificity; leave-one-out cross-validation tests each subject individually.
- Model ensembling and stacking: Ensembling combines model predictions through averaging or voting, whereas stacking feeds classifier outputs into another algorithm for final classification.These approaches may improve accuracy and robustness or reduce overfitting, but can make interpretation less straightforward.
- Model ensembling and stacking: Model selection may favor a slightly less accurate single model when interpretability of disease-related brain regions is more important.Ensembles can sometimes provide feature importance, while random forests combine individual decision trees.
3. Practical issues
Reliable structural-neuroimaging classifiers require careful preprocessing, feature handling, model tuning, and cross-validation. The review emphasizes preventing leakage and balancing predictive performance against interpretability and generalizability.
- Cross-validation: Cross-validation must include internal loops for feature selection and hyperparameter tuning; performing these steps on the full sample produces overly optimistic classifiers.Permutation tests can further enhance evaluation.
- Preprocessing: Feature preprocessing may include centering, scaling, dimensionality reduction, and disease-specific correction of confounding effects.Removing age-related effects from VBM data improved accuracy by approximately 2%, with a later effect of 5% reported for MCI conversion prediction.
- Feature representation: Atlas-based averaging reduces imaging dimensionality but can lose differences when anatomical regions are too large.This trade-off is especially relevant when voxel-level data contain many variables.
- Feature selection: Feature selection can improve generalization and interpretability by removing irrelevant or redundant variables and identifying brain regions associated with classification.SVMs can tolerate irrelevant features, but accuracy decreases relative to using only relevant features.
- Model selection: Model choice depends on dimensionality and overfitting risk: Gaussian SVMs performed better than linear SVMs in lower-dimensional problems, whereas nonlinear models can learn noisy patterns.The review also notes that feature importance can be mapped spatially to support biological interpretation.
- Leakage: Leakage occurs when outcome-informed feature selection uses the entire dataset before cross-validation; test-case labels should be reserved for final evaluation.Such errors can create variables carrying information about the outcome and inflate apparent performance.
4. Machine learning applied to structural neuroimaging
The review focuses on predictive classification studies using machine learning with structural neuroimaging, while largely excluding non-categorical prediction tasks. Accuracy is reported consistently as a value between 0 and 1.
- Scope: The review centers on studies providing predictive classification and generally excludes papers without categorical classification outcomes.A few studies with non-categorical outputs, such as autism ADOS scores, are included as exceptions.
- Evaluation: Classification accuracy is the proportion of correct predictions, calculated from true-positive and true-negative predictions divided by all predictions.Reported accuracy scores use a 0-to-1 scale.
4.1. Alzheimer’s disease/ mild cognitive impairment
Structural-neuroimaging machine learning studies in Alzheimer’s disease and MCI address diagnosis and prognosis, with performance varying by task and methodology. Across reviewed studies, medial temporal and related regions repeatedly support classification, while methodological choices affect robustness and accuracy.
- Classification targets: AD studies commonly classify patients against controls, distinguish AD or controls from MCI, or predict which MCI patients will progress to AD.The first two tasks address diagnosis, whereas MCI-nc versus MCI-c addresses prognosis.
- Task difficulty: AD versus MCI classification is more difficult than AD versus controls because MCI can occupy a diagnostic gray area and resemble mild AD or cognitive impairment.Predicting progression from MCI within a specified time window is more challenging still.
- Multimodal imaging: 0.89 versus 0.86 AUC resulted from adding DTI to structural T1 information, whereas adding rs-fMRI to all three modalities reduced accuracy to 0.79 and AUC to 0.82.The authors attributed the decline to possible noise and SVM overfitting in rs-fMRI data.
- Methodological caveats: Feature-selection leakage affected several studies because variables were selected using the entire dataset rather than within cross-validation.The review therefore cautions that reported kernel comparisons may reflect methodological leakage in some studies.
- Feature selection: Feature selection improved accuracy by up to 5%, and AD-related regions supported MCI-nc versus MCI-c classification with accuracy up to 0.745 and AUC = 0.766.This indicates that regions identified in AD-versus-control classification can also be useful for predicting MCI conversion.
- Reported performance: 0.63 accuracy was obtained on the unobserved CADDementia test dataset by the challenge-winning multi-class LDA classifier.More complex classifiers did not improve performance in that comparison.
- Informative regions: Hippocampus, temporal lobes, amygdala, parahippocampal gyrus, middle temporal gyrus, entorhinal cortex, and insula are repeatedly identified as important gray-matter regions for AD classification.The review reports similar affected regions across AD and MCI cohorts in its regional summary.
4.2. Autism
Structural MRI studies classify autism and related symptom-defined groups using diverse modalities and algorithms, with performance varying by subgroup, hemisphere, and brain region.
- Some studies used ADOS as a continuous clinical score rather than a binary autistic-versus-control label.The section summarizes both binary classification and continuous-score prediction approaches.
- ASD/LI versus ASD/LI+ classification achieved 0.7 accuracy and was described as more challenging than ASD-versus-control classification.The labels targeted language-impairment-related symptom differences within ASD.
- Gray-matter effects were broadly distributed across the brain, with subcortical regions achieving higher accuracies than cortical regions.The review therefore did not provide a table of relevant regions for ASD classification.
4.3. Multiple sclerosis
Structural MRI classification in multiple sclerosis spans diagnosis, subgroup discrimination, and conversion prediction, with accuracy depending on the comparison and prediction horizon.
- The review found relatively few structural-difference classification studies in MS, while machine-learning applications focused more heavily on automatic white-matter-lesion segmentation.The review attributes this emphasis to the direct diagnostic usefulness of detecting white-matter lesions and the challenge of labeling them automatically.
- Lesion information yielded accuracy up to 0.96, compared with 0.84 using normal-appearing gray matter and 0.91 using normal-appearing white matter.These results came from segmenting T1 and T2 images into lesions and normal-appearing tissue regions.
- Healthy-control versus pediatric-MS classification reached 0.87 accuracy, whereas distinguishing early- from late-onset pediatric MS reached 0.807.
- Gray-matter information produced accuracies of 0.85, 0.83, and 0.77 for three MS subgroup comparisons.The comparisons were early versus late MS, low versus high white-matter lesion load, and benign versus non-benign MS, respectively.
- Conversion prediction from clinically isolated syndrome reached 0.714 at 1 year and 0.68 at 3 years, making the longer-horizon task harder.Optimal feature combinations also differed between the 1-year and 3-year prediction tasks.
4.4. Parkinson’s disease and related disorders
Parkinsonian-disorder studies extend beyond binary diagnosis to multiclass classification, with white-matter features often informative but several analyses affected by methodological concerns.
- Multiclass PD-related classification included multinomial-logit models and reported the highest multiclass Brier error score when using gray matter only.
- Several reported results are constrained by leakage, including whole-sample feature weighting, hyperparameter-tuning omissions, and heterogeneous diagnostic categories.The review specifically flags feature processing outside cross-validation and an “Other” category containing MSA and PSP.
- White-matter voxel values produced high control-versus-PSP and PD-versus-PSP accuracies, while gray-matter values yielded much lower accuracies.One related study reported perfect classification using white-matter values alone.
- Including healthy controls in multiclass classification lowered overall accuracy scores.The reviewed multiclass studies included combinations of PD, PSP, MSA, MSA-P, MSA-C, and healthy controls.
- The review concludes that PD, PSP, and MSA affect different brain regions despite similar symptoms.Relevant gray- and white-matter regions are summarized in the paper’s regional table.
4.5. Other
Structural and multimodal imaging classifiers were reviewed across ADHD, depression, schizophrenia, traumatic brain injury, stroke, anorexia nervosa, and bipolar disorder, with performance varying by task and methodological rigor.
- ADHD: ADHD classification accuracies ranged from 0.611 without statistical significance to 0.793 and up to 0.902 across reviewed studies.The studies used T1-derived gray-matter or cortical features with linear SVM, Gaussian-process, or extreme-learning methods.
- Methodological considerations: Feature selection outside cross-validation was reported in several studies, creating leakage concerns for some high-accuracy results.The review also notes a goal of standardized algorithm comparisons to support a future accuracy benchmark.
- Traumatic brain injury: Traumatic-brain-injury classification reached 0.86 with a multilayer perceptron, 0.80 with a Bayesian network, and 0.934 with a linear SVM.These results came from multimodal or DTI-derived graph features, respectively.
- Stroke: For stroke, linear-SVM accuracy was 0.826 for patient-versus-control classification and 0.876 for predicting motor impairment, while DWI-based hand-impairment classification had sensitivity 0.50.The review concludes that resting-state fMRI was more useful than DTI for predicting behavioural deficits.
5. Discussion
The review surveys structural-MRI machine-learning applications across diseases and clinical classification tasks, while highlighting methodological and translational limitations. Performance is strongest for disease-versus-control classification, but weaker for clinically relevant progression predictions and vulnerable to leakage, heterogeneity, and limited validation.
- Scope and contribution: The review compiles machine-learning techniques for structural neuroimaging across cross-validation practices and applications in multiple diseases and disorders.It aims to provide a broad view of the field and its current state of the art.
- Clinical framing: Binary comparisons dominate the literature, although multiclass approaches may be more clinically insightful because binary classifiers require eliminating all but two clinical labels.The review notes that multiclass approaches have been used only in a few cases.
- Classification performance: Accuracies exceeded 0.9 for distinguishing patients from healthy controls in several diseases, including AD, autism, MS, and PD.These results suggest that disease-related brain patterns identifiable on MRI can distinguish diseased populations from controls.
- Classification performance: Accuracies were generally around 0.7 when distinguishing progressive from stable patients, including clinically important intermediate states such as MCI-c versus MCI-nc.Binary approximations remain useful, but many disorders are spectrum disorders whose subtleties may not be fully captured by binary labels.
- Methodological limitations: Feature-selection leakage can produce biased, overly optimistic results, especially in the predominantly small-sample studies reviewed.Leakage occurs when informative variables are selected using the full sample or test-set labels outside the cross-validation loop; nearly all reviewed studies had fewer than 250 samples.
- Methodological limitations: Model generalizability is limited by scanner and acquisition heterogeneity, so estimated accuracies may be reliable mainly for data from similar scanners and parameters.Different MRI contrasts and acquisition parameters can reduce applicability to new data and clinical practice.