Source-linked AI summary
ASD-DiagNet: A hybrid learning approach for detection of Autism Spectrum Disorder using fMRI data
Taban Eslami, Vahid Mirjalili, Alvis Fong, Angela Laird, Fahad Saeed
TL;DR
ASD diagnosis lacks a definitive quantitative biological test, motivating more objective fMRI-based approaches. ASD-DiagNet combines selected connectivity features, joint autoencoder–perceptron learning, and feature-space augmentation; it achieves 70.1% accuracy in 40 minutes on the full dataset.
Problem
ASD is difficult to diagnose because behavioural symptoms overlap and no definitive quantitative biological test exists.
Method
ASD-DiagNet uses selected fMRI connectivity features, jointly trained autoencoder and single layer perceptron networks, and linear-interpolation data augmentation.
Results
70.1% accuracy was achieved through 10-fold cross-validation on 1035 subjects in 40 minutes, compared with 6 hours for the state-of-the-art method.
Takeaways & Limitations
The results support use of the approach for small intra-site datasets and larger multi-site ABIDE datasets within a reasonable runtime.
Abstract
from arXiv · showhide
Mental disorders such as Autism Spectrum Disorders (ASD) are heterogeneous disorders that are notoriously difficult to diagnose, especially in children. The current psychiatric diagnostic process is based purely on the behavioural observation of symptomology (DSM-5/ICD-10) and may be prone to over-prescribing of drugs due to misdiagnosis. In order to move the field towards more quantitative fashion, we need advanced and scalable machine learning infrastructure that will allow us to identify reliable biomarkers of mental health disorders. In this paper, we propose a framework called ASD-DiagNet for classifying subjects with ASD from healthy subjects by using only fMRI data. We designed and implemented a joint learning procedure using an autoencoder and a single layer perceptron which results in improved quality of extracted features and optimized parameters for the model. Further, we designed and implemented a data augmentation strategy, based on linear interpolation on available feature vectors, that allows us to produce synthetic datasets needed for training of machine learning models. The proposed approach is evaluated on a public dataset provided by Autism Brain Imaging Data Exchange including 1035 subjects coming from 17 different brain imaging centers. Our machine learning model outperforms other state of the art methods from 13 imaging centers with increase in classification accuracy up to 20% with maximum accuracy of 80%. The machine learning technique presented in this paper, in addition to yielding better quality, gives enormous advantages in terms of execution time (40 minutes vs. 6 hours on other methods). The implemented code is available as GPL license on GitHub portal of our lab (https://github.com/pcdslab/ASD-DiagNet).
I. INTRODUCTION
ASD is difficult to diagnose quantitatively because behavioural symptoms overlap and no definitive biological test exists. The paper introduces ASD-DiagNet, an fMRI-only classifier intended to improve accuracy, scalability, and diagnostic quantification.
- ASD diagnosis is difficult because the disorder is heterogeneous, symptoms overlap, and current assessment relies on behavioural observation.
- Quantitative brain-imaging analysis and machine learning are investigated as sources of biomarkers for diagnosing brain disorders, including ASD.
- ASD-DiagNet combines an autoencoder, a single layer perceptron, and feature-space data augmentation to classify ASD subjects from healthy controls using fMRI.
- 70.1% classification accuracy was achieved, outperforming the cited state-of-the-art technique; average individual-site accuracy was 63%, 7% above the reported comparison.
- 41 minutes of execution time was reported for ASD-DiagNet versus 6 hours for other methods, while the study evaluated all 1035 ABIDE subjects across 17 sites.
III. MATERIALS AND METHODS
The study uses preprocessed resting-state fMRI from ABIDE-I, collected across multiple imaging sites and represented through atlas-based regional time series.
- ABIDE-I contains 1112 resting-state fMRI datasets collected from 17 sites, and the study analyzes 505 ASD subjects and 530 healthy controls.
- The data provide average time series from regions of interest defined by seven atlases and four preprocessing pipelines.
B. ASD-DiagNet: Feature extraction and classification
ASD-DiagNet converts selected functional-connectivity correlations into compact features, then jointly trains an autoencoder and perceptron for reconstruction and classification.
- Pearson correlations between regional fMRI time series provide functional-connectivity features for classification.
- For the 200-region CC-200 atlas, 19,900 pairwise correlations are reduced to 9,950 selected high- and anti-correlations.
- The autoencoder maps input features to a lower-dimensional bottleneck representation and reconstructs the original input through its decoder.
- The single layer perceptron uses bottleneck features to estimate ASD probability with sigmoid activation and binary cross-entropy training.
- Unlike separate training, ASD-DiagNet jointly optimizes reconstruction and classification losses so extracted features support both objectives.
C. Data augmentation using linear interpolation
The proposed augmentation method synthesizes feature vectors by interpolating each training sample with a randomly selected same-class neighbor, using EROS to identify neighbors. One synthetic sample is generated per training point, doubling the training set, while using only the first two eigenvectors reduces similarity-computation time without affecting results.
- Method: SMOTE-inspired augmentation generates synthetic feature vectors from a sample and one of its nearest neighbors.The neighbor is selected after computing similarities in feature space.
- Neighbor selection: EROS measures similarity between multivariate time series using covariance-matrix eigenvalues and corresponding eigenvectors.The feature vectors contain 9950 values, making Euclidean pairwise distances inefficient.
- Neighbor selection: EROS weights eigenvector similarities using weights derived by normalizing eigenvalues across the training dataset.The weight vector is computed by aggregating normalized eigenvalues with the mean function.
- Efficiency: Using only the first two eigenvectors significantly reduces runtime without affecting experimental results.This simplification replaces the use of all eigenvectors and eigenvalues for pairwise similarity computation.
- Augmentation procedure: Each training sample receives one synthetic sample formed by linear interpolation with a randomly selected one of its 5 nearest neighbors, doubling the training set.The neighbor search uses EROS, and the interpolation coefficient is sampled from [0.5, 1].
IV. EXPERIMENTS AND RESULTS
The experiments evaluate ASD-DiagNet with EROS-based feature-space augmentation across the complete heterogeneous dataset and separately within each imaging site. The augmentation algorithm finds five nearest neighbors, selects one randomly, and interpolates a synthetic sample for every training point.
- Experimental setup: The evaluation uses a Linux server with two Intel Xeon E5-2620 processors, 48 GB of RAM, and an NVIDIA Tesla K-40c GPU.CUDA 8 and PyTorch were used for the experiments.
- Data augmentation: Figure 2 depicts augmentation by selecting a sample, choosing a same-class nearest neighbor, and linearly interpolating between them.The procedure generates a new sample p′ from p and qr.
- Data augmentation: The augmentation algorithm finds 5 nearest neighbors using EROS, randomly selects one, and creates one interpolated sample for each training point.The synthetic sample uses a random coefficient in [0.5, 1] and is appended to the training set.
- Experimental design: Experiments assess ASD-DiagNet on all 1035 subjects and separately within each data center.The whole-dataset phase tests generalization across heterogeneous scanning parameters, whereas site-specific evaluation addresses intra-site training.
A. Phase 1: Experiments using the whole dataset
In whole-dataset evaluation, ASD-DiagNet is compared with Heinsfeld et al., random forest, and an RBF-kernel SVM using 10-fold cross-validation on 1035 subjects. It achieves the highest reported accuracy, with augmentation improving results by around 1%.
- Whole-dataset evaluation: 70.1% accuracy is achieved by ASD-DiagNet with 10-fold cross-validation on the whole 1035-subject dataset.The comparison includes Heinsfeld et al., random forest, and an RBF-kernel SVM.
- Whole-dataset evaluation: ASD-DiagNet with data augmentation achieves the highest accuracy among the existing methods in Table II.The table reports classification performance using 10-fold cross-validation on the whole dataset.
B. Phase 2: Intra-site evaluation
In the intra-site evaluation, ASD-DiagNet is assessed with 5-fold cross-validation separately for each data center. It achieves the highest average accuracy among the compared methods, while augmentation improves accuracy overall and especially at OHSU.
- Intra-site evaluation: ASD-DiagNet achieves the highest accuracy in most individual-site cases and outperforms other methods on average.Each data center is evaluated separately using 5-fold cross-validation.
- Data augmentation: Around 2% overall accuracy improvement comes from the proposed data augmentation in the intra-site evaluation.At OHSU, augmentation produces a 15% increase in accuracy.
C. Running time
ASD-DiagNet was evaluated for running time against other approaches using 10-fold cross-validation on the whole dataset. The method was reported as significantly faster, while its augmentation procedure doubled training data and doubled computation time.
- ASD-DiagNet performs significantly faster than the compared approach.The comparison concerns training and evaluation for 10-fold cross-validation on the whole dataset.
- One artificial sample per training subject doubles the training-set size.The synthetic samples are generated through the proposed data augmentation process.
- A factor-of-2 increase in computation time accompanies the data augmentation.
D. Experiment on other parcellations
The study tested ASD-DiagNet with AAL and Dosenbach160 brain parcellations, selecting the largest and smallest correlations as input features. With both atlases, the proposed method outperformed existing methods, although accuracy was below that obtained with CC-200.
- Atlas configurations: AAL parcellates the brain into 116 regions, whereas Dosenbach160 parcellates it into 160 regions.
- Atlas configurations: 12,720 and 6,670 pairwise correlations are available for Dosenbach160 and AAL, respectively.
- Feature selection: Half of the correlations were selected as input features by retaining the 1/4 largest and 1/4 smallest values.
- Results: ASD-DiagNet with and without augmentation performs better than existing methods using the alternative parcellations.
- Results: Classification accuracy using AAL and Dosenbach160 is below the accuracy obtained using CC-200.The authors attribute this to more discriminative pairwise-correlation patterns among CC-200 regions.
V. CONCLUSION AND FUTURE WORK
The paper classifies ASD versus healthy controls from multi-center ABIDE fMRI data without demographic information, using selected brain connections, joint representation learning, and feature-space augmentation. Across whole-dataset and individual-center evaluations, it reports higher accuracy than the state-of-the-art method with substantially shorter runtime.
- ASD-DiagNet classifies ASD subjects and healthy controls using selected correlated and anti-correlated brain connections from fMRI data.
- The autoencoder and single layer perceptron are jointly trained for feature selection and classification.
- 70.1% accuracy was achieved in 40 minutes with 10-fold cross-validation on the whole dataset.
- The approach required substantially less time than the 6 hours needed by the state-of-the-art method while achieving higher classification accuracy.
- Data augmentation improved accuracy by around 2% in the separate 5-fold cross-validation experiments conducted for each data center.
FUNDING
The research was supported by NIGMS and NSF awards. The authors state that the content is solely their responsibility and may not represent governmental agencies’ official views.
- NIGMS supported the research through NIH Award Number R15GM120820.
- NSF supported the research through awards NSF CRII CCF-1464268, NSF CRII CCF-1855441, NSF CAREER ACI-1651724, and NSF OAC 1925960.
- The authors state that the content is solely their responsibility and does not necessarily represent governmental agencies’ official views.