Outlier-based Autism Detection using Longitudinal Structural MRI

Devika K, Venkata Ramana Murthy Oruganti, Dwarikanath Mahapatra, Ramanathan SubramanianarXiv 2022 · DOI · arXiv

TL;DR

This paper introduces an unsupervised outlier-detection framework for Autism Spectrum Disorder (ASD) using longitudinal structural MRI (sMRI) scans. By training a Self-Attention Generative Adversarial Network (SAGAN) exclusively on healthy subjects to predict sequential brain slices, the system flags ASD scans as "outliers" based on high reconstruction errors. This approach achieves state-of-the-art accuracy while requiring significantly less training data than traditional supervised methods.

What Problem Does This Paper Address?

Diagnosing Autism Spectrum Disorder (ASD) typically relies on clinical behavioral evaluations and cognitive tests, which can be highly subjective and prone to clinical bias. While machine learning applied to structural Magnetic Resonance Imaging (sMRI) offers an objective alternative, current approaches face two major hurdles:

  1. Data Scarcity: Most deep learning models require large, supervised datasets containing many labeled ASD examples. Because medical data is scarce and difficult to collect, training these models is challenging.
  2. Static Snapshots: The majority of neuroimaging studies rely on cross-sectional data (a single scan per subject). This completely misses the spatio-temporal changes in brain connectivity that occur as a subject ages.

The authors address these gaps by framing ASD detection as an unsupervised anomaly (or outlier) detection problem using longitudinal data (multiple scans from the same individual over time).

Methods

Instead of training a classifier on both healthy and ASD scans, the authors developed an unsupervised framework that learns only what a healthy brain looks like.

The core task is multiple sMRI slice reconstruction. The model is fed a stack of three consecutive sMRI slices and tasked with generating the next three adjacent slices.

Figure 1: Approach overview: Longitudinal sMRI scans are pairwise-aligned via the Freesurfer pipeline. Slices along a given plane (Axial, Coronal and Sagittal) are extracted and three adjacent slices stacked. Different architectures (UNet, GAN, SAGAN) are trained using only sMRI scans of healthy subjects. Reconstruction error from training is used to compute a threshold for outlier direction. A test subject is labeled as ASD if reconstruction error is greater than the threshold.

The authors evaluated three different encoder-decoder architectures for this task: a standard UNet, a conventional Generative Adversarial Network (GAN), and a Self-Attention GAN (SAGAN).

During inference, if the model encounters an sMRI scan from a patient with ASD, its structural deviations (such as enlarged ventricles) cause the reconstruction to fail. This results in a high reconstruction loss, safely flagging the scan as an outlier (ASD).

To optimize performance, the authors combined standard $L_2$ pixel loss with a cosine similarity loss metric, which forces the network to preserve both structural integrity and fine-grained visual features.

Datasets

The framework was evaluated using the ABIDE II (Autism Brain Imaging Data Exchange) dataset, specifically focusing on the longitudinal subsets from UCLA and UPSM.

  • Training Set: 24 scans from 12 healthy subjects (2 scans per subject over a 1-2 year period).
  • Testing Set: 6 scans from 3 healthy subjects and 40 scans from 20 ASD subjects.

To demonstrate the superiority of longitudinal data, the authors also benchmarked their model against cross-sectional data sampled from the ABIDE I dataset.

Key Insights and Ideas

  • The Power of Longitudinal Data: Models trained on longitudinal data achieved a 17% to 28% higher accuracy than identical architectures trained on cross-sectional data. Tracking how brain structures evolve over time provides crucial diagnostic signals that a single snapshot misses.
  • Coronal Plane Superiority: Brain scans can be sliced along three planes: Axial, Coronal, and Sagittal. The experiments revealed that the Coronal plane best encodes the structural connectivity changes associated with ASD, outperforming the Sagittal plane by roughly 20%.
  • Global Context via Self-Attention: Standard convolutional layers are limited by local receptive fields, often yielding blurry reconstructions. Incorporating Self-Attention modules (SAGAN) allowed the network to capture long-range spatial dependencies across the brain, significantly sharpening the generated slices and boosting anomaly detection sensitivity.

Key Results

  • Top Performance: The SAGAN model utilizing a combined Axial and Coronal multi-modality input achieved the highest overall detection accuracy of 95.65% with an Area Under the Curve (AUC) of 0.90.
  • Data Efficiency: The unsupervised framework achieved a 95.65% accuracy using 20 times fewer training samples than the leading state-of-the-art supervised method (which scored 96.60% but required hundreds of labeled scans).

Limitations and Future Work

The primary limitation acknowledged by the authors is the small sample size of the longitudinal dataset available in ABIDE II (only 38 subjects total after preprocessing).

For future work, the authors propose:

  1. Extending this unsupervised outlier framework to detect other neurological conditions, such as Attention Deficit Hyperactivity Disorder (ADHD) and Schizophrenia.
  2. Utilizing multi-task learning to simultaneously exploit large, readily available cross-sectional datasets alongside sparse, high-value longitudinal data.
  3. Exploring alternative deep architectures, such as Dense-Attentive GANs.

Summary updated

Abstract

Diagnosis of Autism Spectrum Disorder (ASD) using clinical evaluation (cognitive tests) is challenging due to wide variations amongst individuals. Since no effective treatment exists, prompt and reliable ASD diagnosis can enable the effective preparation of treatment regimens. This paper proposes structural Magnetic Resonance Imaging (sMRI)-based ASD diagnosis via an outlier detection approach. To learn Spatio-temporal patterns in structural brain connectivity, a Generative Adversarial Network (GAN) is trained exclusively with sMRI scans of healthy subjects. Given a stack of three adjacent slices as input, the GAN generator reconstructs the next three adjacent slices; the GAN discriminator then identifies ASD sMRI scan reconstructions as outliers. This model is compared against two other baselines -- a simpler UNet and a sophisticated Self-Attention GAN. Axial, Coronal, and Sagittal sMRI slices from the multi-site ABIDE II dataset are used for evaluation. Extensive experiments reveal that our ASD detection framework performs comparably with the state-of-the-art with far fewer training data. Furthermore, longitudinal data (two scans per subject over time) achieve 17-28% higher accuracy than cross-sectional data (one scan per subject). Among other findings, metrics employed for model training as well as reconstruction loss computation impact detection performance, and the coronal modality is found to best encode structural information for ASD detection.
BibTeX
@misc{k2022outlier,
  author = {Devika K and Venkata Ramana Murthy Oruganti and Dwarikanath Mahapatra and Ramanathan Subramanian},
  title = {{Outlier-based Autism Detection using Longitudinal Structural MRI}},
  year = {2022},
  eprint = {2202.09988},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2202.09988},
}
Figure