[Paper Review] Unsupervised Detection of Lesions in Brain MRI using constrained adversarial auto-encoders

Xiaoran Chen, Ender Konukoglu|arXiv (Cornell University)|Jun 13, 2018

Brain Tumor Detection and ClassificationNeuroscience9 references122 citations

TL;DR

The paper proposes an unsupervised lesion detection method in brain MRI by learning healthy brain distributions with VAE/AAE, introducing a latent-space consistency constraint to improve lesion detection, and evaluating on BRATS with HCP healthy data.

ABSTRACT

Lesion detection in brain Magnetic Resonance Images (MRI) remains a challenging task. State-of-the-art approaches are mostly based on supervised learning making use of large annotated datasets. Human beings, on the other hand, even non-experts, can detect most abnormal lesions after seeing a handful of healthy brain images. Replicating this capability of using prior information on the appearance of healthy brain structure to detect lesions can help computers achieve human level abnormality detection, specifically reducing the need for numerous labeled examples and bettering generalization of previously unseen lesions. To this end, we study detection of lesion regions in an unsupervised manner by learning data distribution of brain MRI of healthy subjects using auto-encoder based methods. We hypothesize that one of the main limitations of the current models is the lack of consistency in latent representation. We propose a simple yet effective constraint that helps mapping of an image bearing lesion close to its corresponding healthy image in the latent space. We use the Human Connectome Project dataset to learn distribution of healthy-appearing brain MRI and report improved detection, in terms of AUC, of the lesions in the BRATS challenge dataset.

Motivation & Objective

• Motivate unsupervised lesion detection by leveraging prior healthy brain appearance to reduce labeled data needs.
• Learn a healthy brain distribution using auto-encoder based models (VAE/AAE).
• Address latent-space consistency to improve separation between healthy and abnormal images.
• Demonstrate improved lesion detection on BRATS using healthy-distribution models trained on HCP data.

Proposed method

• Train VAE and AAE on healthy brain T2-weighted MRIs from the Human Connectome Project.
• Replace standard divergence with WGAN-GP for AAE to better match latent distributions.
• Introduce a latent-space consistency constraint by adding a regularization term ||z_h - z_h' ||^2 to push healthy and reconstructed latent codes closer.
• Use pixel-wise reconstruction residuals |X_a - X_a'| to detect lesions in BRATS data.
• Down-sample images to 32x32 for feasible reconstruction quality and training stability.

Experimental results

Research questions

• RQ1 Can unsupervised models learned from healthy brain images detect brain lesions without labeled lesion data?
• RQ2 Does enforcing latent-space consistency improve lesion detection compared to standard VAE/AAE?
• RQ3 Which model and parameter settings yield best pixel-wise anomaly detection (AUC) on BRATS using healthy-training distributions?
• RQ4 How do reconstructions differ for healthy versus abnormal regions and how does this affect residual-based lesion detection?

Key findings

• AAE with latent constraint (λ = 1.0) achieved the highest AUC of 0.923 on BRATS BRATS-2015 T2 data.
• Latent constraint improves reconstruction realism in healthy regions and highlights lesion areas more accurately in residuals.
• λ = 1.0 provided clearer separation between healthy and anomalous pixel reconstruction-error distributions, reducing overlap to 17% compared to higher overlaps for other models.
• Compared to VAE and AAE with no constraint, the constrained AAE yielded sharper reconstructions and more accurate lesion localization in residuals.