[Paper Review] Comparison of Probabilistic Deep Learning Methods for Autism Detection

Godfrin Ismail, Kenneth Chesoli+2 more|arXiv (Cornell University)|Mar 9, 2023

Autism Spectrum Disorder Research4 citations

TL;DR

This study evaluates probabilistic deep learning models for autism spectrum disorder (ASD) detection using multimodal data—fMRI, MRI, and facial images—demonstrating that transfer learning models like Xception and VGG16 achieve high accuracy (up to 95%) in classifying ASD, with Xception reaching 91% on facial data and VGG16 showing 95% accuracy on fMRI data, highlighting their potential for early, automated diagnosis.

ABSTRACT

Autism Spectrum Disorder (ASD) is one neuro developmental disorder that is now widespread in the world. ASD persists throughout the life of an individual, impacting the way they behave and communicate, resulting to notable deficits consisting of social life retardation, repeated behavioural traits and a restriction in their interests. Early detection of the disorder helps in the onset treatment and helps one to lead a normal life. There are clinical approaches used in detection of autism, relying on behavioural data and in worst cases, neuroimaging. Quantitative methods involving machine learning have been studied and developed to overcome issues with clinical approaches. These quantitative methods rely on machine learning, with some complex methods based on deep learning developed to accelerate detection and diagnosis of ASD. These literature is aimed at exploring most state-of-the-art probabilistic methods in use today, characterizing them with the type of dataset they're most applied on, their accuracy according to their novel research and how well they are suited in ASD classification. The findings will purposely serve as a benchmark in selection of the model to use when performing ASD detection.

Motivation & Objective

• To evaluate state-of-the-art probabilistic deep learning models for autism spectrum disorder (ASD) detection across diverse data modalities.
• To identify the most accurate and robust deep learning architectures for ASD classification based on performance metrics and data type.
• To provide a benchmark for selecting optimal models depending on available data—fMRI, MRI, or facial images—supporting early diagnosis.
• To assess the impact of data augmentation and hyperparameter tuning on model generalization and accuracy.
• To guide clinicians and researchers in adopting deep learning for scalable, early ASD detection with high reliability.

Proposed method

• Trained and compared multiple deep learning models: Xception, VGG19, NASNetMobile, Auto-ASD-Network, VGG16, and a custom DNN on ASD detection tasks.
• Utilized transfer learning with pre-trained models (e.g., Xception, VGG19) fine-tuned on facial images from the Kaggle dataset to classify autism vs. typical development.
• Applied data augmentation via SMOTE to address class imbalance in fMRI and MRI datasets, improving model generalization and reducing overfitting.
• Employed a two-stage classification pipeline: first, a deep neural network extracted features; second, an SVM classifier with Auto-Tune Models (ATM) optimized hyperparameters (kernel, penalty) for improved performance.
• Evaluated models using multiple metrics: accuracy, sensitivity, specificity, and confusion matrices to ensure robust performance assessment.
• Used the ABIDE fMRI dataset for model training and validation, with feature extraction from hidden layers to assess representation quality.

Figure 1: Class information in the ABIDE Dataset

Experimental results

Research questions

• RQ1 How do probabilistic deep learning models perform in classifying autism spectrum disorder across different data modalities (fMRI, MRI, facial images)?
• RQ2 Which deep learning architectures achieve the highest classification accuracy for ASD detection, and how do they compare in terms of sensitivity and specificity?
• RQ3 To what extent does data augmentation (e.g., SMOTE) and hyperparameter tuning (e.g., via ATM) improve model performance in ASD classification?
• RQ4 How effective are transfer learning models like Xception and VGG19 in detecting ASD from facial features compared to traditional deep neural networks?
• RQ5 Which model is most suitable for early ASD detection based on data availability and performance metrics?

Key findings

• VGG16 achieved the highest classification accuracy of 95% on fMRI data, outperforming other models in the study.
• The Xception model achieved 91% testing accuracy on facial image data, with 132 true positives and 18 false negatives among 150 autistic children.
• The Auto-ASD-Network model reached 80% accuracy on four fMRI datasets, improving SVM performance by 26% and state-of-the-art methods by 14%.
• VGG19 showed a training accuracy increase from 56% to 85% over 25 epochs, followed by a drop to 82% in validation, indicating potential overfitting.
• NASNetMobile performed poorly, with validation accuracy ranging from 65% to 78%, making it the least effective among the three models tested on facial data.
• The combination of SMOTE data augmentation and SVM fine-tuning significantly enhanced model performance, particularly in imbalanced datasets.

Figure 2: The Architecture of VGG16 [ 26 ]