[Paper Review] A novel machine learning based framework for detection of Autism Spectrum Disorder (ASD)

Hamza Sharif, Rizwan Ahmed Khan|arXiv (Cornell University)|Mar 27, 2019

Autism Spectrum Disorder ResearchNeuroscience103 references75 citations

TL;DR

This paper proposes a machine learning framework for automated Autism Spectrum Disorder (ASD) detection using T1-weighted MRI scans from the ABIDE-I dataset. It extracts features from the corpus callosum and intracranial brain volume, applies feature weighting to enhance discriminative power, and evaluates both conventional machine learning and transfer learning with VGG16, achieving 66% accuracy in ASD detection—demonstrating the feasibility of hybrid approaches for neuroimaging-based diagnosis with limited data.

ABSTRACT

Computer vision and machine learning are the linchpin of field of automation. The medicine industry has adopted numerous methods to discover the root causes of many diseases in order to automate detection process. But, the biomarkers of Autism Spectrum Disorder (ASD) are still unknown, let alone automating its detection. Studies from the neuroscience domain highlighted the fact that corpus callosum and intracranial brain volume holds significant information for detection of ASD. Such results and studies are not tested and verified by scientists working in the domain of computer vision / machine learning. Thus, in this study we have proposed a machine learning based framework for automatic detection of ASD using features extracted from corpus callosum and intracranial brain volume from ABIDE dataset. Corpus callosum and intracranial brain volume data is obtained from T1-weighted MRI scans. Our proposed framework first calculates weights of features extracted from Corpus callosum and intracranial brain volume data. This step ensures to utilize discriminative capabilities of only those features that will help in robust recognition of ASD. Then, conventional machine learning algorithm (conventional refers to algorithms other than deep learning) is applied on features that are most significant in terms of discriminative capabilities for recognition of ASD. Finally, for benchmarking and to verify potential of deep learning on analyzing neuroimaging data i.e. T1-weighted MRI scans, we have done experiment with state of the art deep learning architecture i.e. VGG16 . We have used transfer learning approach to use already trained VGG16 model for detection of ASD. This is done to help readers understand benefits and bottlenecks of using deep learning approach for analyzing neuroimaging data which is difficult to record in large enough quantity for deep learning.

Motivation & Objective

• To develop a machine learning framework for automated ASD detection using neuroimaging data.
• To investigate the discriminative potential of corpus callosum and intracranial brain volume features in ASD classification.
• To compare conventional machine learning with deep learning (VGG16) on limited neuroimaging data for ASD detection.
• To evaluate the benefits and limitations of deep learning in low-data regimes for neuroimaging applications.
• To support early clinical detection of ASD by leveraging structural MRI features with robust feature selection.

Proposed method

• Extracts features from corpus callosum and intracranial brain volume regions in T1-weighted MRI scans from the ABIDE-I dataset.
• Applies feature weighting to prioritize discriminative features for improved classification robustness.
• Employs conventional machine learning algorithms on selected features to classify ASD vs. control subjects.
• Utilizes transfer learning with a pre-trained VGG16 model fine-tuned on cropped tif images of the corpus callosum.
• Trains the final dense layer of VGG16 using ADAM optimizer and softmax activation on 80% of the data, validating on 20%.
• Processes nifti-formatted MRI scans into TIFF format, discards non-representative slices, and retains ~100k images with complete corpus callosum regions.

Experimental results

Research questions

• RQ1 Can features from the corpus callosum and intracranial brain volume improve ASD detection accuracy in structural MRI?
• RQ2 How does feature weighting enhance the performance of conventional machine learning models in ASD classification?
• RQ3 What is the performance of transfer learning with VGG16 on limited neuroimaging data for ASD detection?
• RQ4 How do conventional machine learning and deep learning compare in terms of accuracy and data efficiency for ASD detection?
• RQ5 What are the key bottlenecks of using deep learning for neuroimaging data when labeled data is scarce?

Key findings

• The proposed framework achieved a detection accuracy of 66% using transfer learning with VGG16 on ABIDE-I MRI data.
• Conventional machine learning methods outperformed baseline models but fell short of the 10% improvement seen with deep learning.
• Feature weighting significantly improved classifier robustness and reduced computational complexity.
• The study confirms that corpus callosum and intracranial brain volume features carry discriminative information for ASD classification.
• Transfer learning with VGG16 demonstrated promise despite limited data, though performance was constrained by data scarcity.
• The authors observe a decline in classification accuracy when using multi-site neuroimaging data, consistent with prior findings in the literature.