[Paper Review] Tractometry-based Anomaly Detection for Single-subject White Matter Analysis

Maxime Chamberland, Sila Genc+7 more|arXiv (Cornell University)|May 22, 2020

Advanced Neuroimaging Techniques and Applications12 references4 citations

TL;DR

This paper proposes a tractometry-based deep autoencoder framework for unsupervised anomaly detection in single-subject diffusion MRI, learning normative white matter microstructure from healthy controls to identify deviations in children with rare copy number variants. The method outperformed z-score and PCA-based approaches (AUC: 0.86 vs. 0.53 and 0.61), enabling personalized detection of microstructural abnormalities in clinically heterogeneous and rare cases.

ABSTRACT

There is an urgent need for a paradigm shift from group-wise comparisons to individual diagnosis in diffusion MRI (dMRI) to enable the analysis of rare cases and clinically-heterogeneous groups. Deep autoencoders have shown great potential to detect anomalies in neuroimaging data. We present a framework that operates on the manifold of white matter (WM) pathways to learn normative microstructural features, and discriminate those at genetic risk from controls in a paediatric population.

Motivation & Objective

• Address the lack of single-subject analysis frameworks in diffusion MRI, especially for rare or clinically heterogeneous cases.
• Enable individualized diagnosis by detecting microstructural deviations from normative patterns without requiring labeled patient data.
• Overcome limitations of voxel-based normative modeling by operating on the anatomically meaningful manifold of white matter tracts.
• Develop an unsupervised framework that learns normative microstructural features from healthy controls only, for subsequent outlier detection in unseen subjects.
• Facilitate transition from group-level comparisons to personalized medicine in neuroimaging by identifying subject-specific microstructural anomalies.

Proposed method

• Acquired high b-value diffusion MRI data (b = 6000 s/mm²) from 90 typically developing children and 8 with copy number variants.
• Extracted rotationally invariant spherical harmonic (RISH0) features from the highest b-value shell for each voxel in 20 white matter tracts.
• Applied tractometry to sample 20 locations per tract, concatenating profiles into a 400-dimensional feature vector per subject.
• Trained a symmetric deep autoencoder with five fully connected layers, using ReLU activations and mean squared error loss to minimize reconstruction error.
• Used a held-out validation set (16 subjects, including 8 CNV and 8 TD) to compute anomaly scores as mean absolute reconstruction error.
• Repeated training 50 times with bootstrapped data to estimate variability and derive mean anomaly scores, comparing performance to z-score and PCA-based Mahalanobis distance.

Experimental results

Research questions

• RQ1 Can an unsupervised deep autoencoder trained only on healthy controls detect microstructural anomalies in single subjects with rare genetic variants?
• RQ2 How does tractometry-based anomaly detection compare to traditional univariate z-score and multivariate PCA-based outlier detection in identifying CNV subjects?
• RQ3 To what extent can the framework identify tract-specific microstructural deviations that are clinically relevant and statistically significant?
• RQ4 Can this approach generalize to extremely rare cases where group comparisons are infeasible due to small sample sizes?
• RQ5 Does the non-linear modeling capacity of autoencoders improve anomaly detection in high-dimensional dMRI data compared to linear methods?

Key findings

• The deep autoencoder achieved an AUC of 0.86 ± 0.06 in detecting CNV subjects as outliers, significantly outperforming the univariate z-score (AUC: 0.53 ± 0.06) and PCA-based Mahalanobis distance (AUC: 0.61 ± 0.09).
• Reconstructed features of CNV subjects showed significant discrepancies along major association pathways, with statistically identifiable deviations (p < 0.01) in multiple tracts.
• In contrast, a representative typically developing subject showed no significant reconstruction errors, confirming the model’s sensitivity to true anomalies.
• The framework successfully identified subject-specific microstructural deviations without requiring prior labels or group-level comparisons.
• The method demonstrated robustness to variability in the training set, with consistent performance across 50 repeated training iterations.
• The results suggest that non-linear, deep autoencoders are better suited than linear methods for detecting subtle, high-dimensional microstructural anomalies in dMRI.