Source-linked AI summary
Unsupervised Detection of Lesions in Brain MRI using constrained adversarial auto-encoders
Xiaoran Chen, Ender Konukoglu
TL;DR
Brain-MRI lesion detection is commonly supervised and dependent on large annotated datasets, whereas humans can detect abnormalities from limited healthy-brain exposure. This paper learns healthy MRI distributions with VAE and AAE models, adds a latent-consistency constraint, and evaluates unsupervised detection on BRATS lesions. The constrained AAE achieved the highest AUC among the evaluated models, while reconstruction quality was limited by down-sampling to 32×32 images.
Problem
Brain-MRI lesion detection remains challenging because supervised approaches require large annotated datasets, despite humans detecting abnormal lesions after seeing a handful of healthy images.
Method
The paper trains VAE and AAE models on healthy MRI distributions and adds a training constraint that encourages lesion-bearing images to map near corresponding healthy images in latent space.
Results
AAE with λ = 1.0 achieved the highest Area Under Curve (AUC) among the models evaluated for lesion detection on BRATS T2w images.
Takeaways & Limitations
The investigated auto-encoder models can detect abnormal lesions using residual images, and the proposed latent constraint improves detection accuracy.
Takeaways & Limitations
Original images were down-sampled to 32×32 because the work does not address methods for reconstructing large high-resolution images.
Abstract
from arXiv · showhide
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.
1 Introduction
Brain-lesion detection is clinically important but increasingly difficult to scale because expert capacity and annotated data lag behind growing imaging demands. The paper therefore studies unsupervised detection by learning healthy-brain appearance and enforcing latent consistency in auto-encoder models.
- Early lesion detection supports prognosis and symptom prevention, while MRI provides in-vivo observations for identifying abnormalities.
- Supervised CNN methods achieve strong segmentation but require many high-quality annotations because their large parameter counts create over-fitting risks.
- Humans can detect abnormal-looking regions after seeing only a few healthy brain images, without extensive training.
- Learning healthy-tissue priors could reduce labeled-data requirements and improve generalization to previously unseen lesions.
- The paper investigates VAE and AAE models trained on healthy HCP MRI, then applies their learned distributions to unsupervised lesion detection in BRATS MRI.
2 Related works
Generative models based on GANs and VAEs approximate high-dimensional image distributions, and related auto-encoder methods use reconstruction error to detect abnormal regions. Prior work includes anomaly detection in retinal OCT and head CT, motivating the paper’s VAE-based comparison.
- GAN- and VAE-based methods are the two main groups of deep generative approaches for approximating high-dimensional imaging distributions.
- Both GAN- and VAE-based methods have been applied to abnormality detection across motion, temporal, and medical imaging data.
- AnoGAN detected abnormalities in healthy retinal optical coherence tomography, while another method used 3D convolutional auto-encoders for abnormal head-CT regions.
- These medical-image methods use reconstruction error as the main abnormality measure, making the approach similar to a basic VAE-based detector.
3 Methodology
The method learns healthy brain-MRI distributions with VAE and AAE models, then detects lesions through reconstruction residuals. A latent-consistency regularizer addresses the possibility that abnormal and healthy images overlap in latent space.
- Generative models: Healthy images are encoded into a lower-dimensional latent representation and reconstructed through learned decoder mappings.The models use a predetermined prior distribution over latent variables; the encoder maps images to z and the decoder reconstructs images from z.
- Anomaly detection: Lesions are detected by reconstructing abnormal images with models trained on healthy images and measuring pixel-wise residual intensity differences.Poor reconstruction indicates low probability under the healthy-image distribution, while the residual highlights abnormal regions.
- Generative models: VAE regularizes latent inference with KL divergence, whereas AAE uses adversarial learning to minimize JS divergence between aggregated posterior and prior.AAE replaces VAE’s KL-divergence mechanism with a GAN-based discriminator-generator objective.
- Latent-space limitation: Healthy-brain variability can exceed lesion-induced differences, so abnormal images may remain close to healthy images in latent space.The paper attributes this issue partly to intensity variation across different slices of 3D brain volumes.
- Representation consistency: The proposed constraint adds a penalty on the distance between latent representations of healthy images and their reconstructions.The regularization weight λ controls the trade-off between the original auto-encoder objective and latent-space consistency.
4 Datasets
The study trains on healthy HCP T2-weighted images and tests lesion detection on BRATS 2015 T2-weighted images. Preprocessing aligns intensity distributions, standardizes images, and downsamples them for reconstruction.
- Datasets: Training uses T2-weighted structural images from 35 healthy HCP subjects, while testing uses T2-weighted images from 42 BRATS 2015 subjects.Lesion detection is performed on the BRATS test images.
- Preprocessing: BRATS images undergo bias correction using n4ITK.The correction is applied specifically to the BRATS challenge dataset.
- Preprocessing: HCP and BRATS histograms are normalized to the same subject-level profile, and images are standardized to zero mean and unit variance.These steps are applied to both datasets unless otherwise specified.
- Resolution constraint: Original images are downsampled to 32×32 because reconstructing large, high-resolution images is challenging for the models.The paper does not address methods for improving reconstruction quality.
5 Results
The models were trained on healthy HCP brain MRI and evaluated for unsupervised lesion detection on BRATS images. Adding latent constraints improved reconstruction consistency and detection, with AAE λ = 1.0 performing best among the tested models.
- Model comparison: AAE was evaluated with λ = 0.5 and λ = 1, alongside VAE and unconstrained AAE models.The comparison used AAE because of its theoretically better latent-space behavior.
- Healthy-image reconstruction: Healthy HCP images were reconstructed to assess how well the models captured the healthy data distribution.Training minimized reconstruction objectives and the models’ respective divergences.
- Lesion detection: BRATS lesions were detected from pixel-wise absolute-intensity residuals between anomalous inputs and their reconstructions.Figure 3 compares original images, reconstructions, residuals, and ground-truth lesions across the four models.
- Improved detection with latent constraint: Latent constraints improved reconstruction and detection, producing healthier reconstructions and residuals that more specifically highlighted lesion areas.Without the constraint, differences extended beyond lesions; constrained AAE reconstructions better preserved realistic brain appearance.
- Improved detection with latent constraint: AAE with λ = 1.0 detected lesions more accurately than the other models, whereas λ = 0.5 produced more false positives and some unrealistic reconstructions.The choice of λ was dataset-dependent in this evaluation.
- Quantitative evaluation: AAE with λ = 1.0 achieved the highest AUC in ROC analysis of pixel-wise detection on BRATS T2-weighted images.The ROC curves compared residual images with ground-truth segmentations.
- Reconstruction-error distributions: The reconstruction-error distributions showed greater healthy-lesion separation for λ = 1, and AAE λ = 1.0 had the least overlap among models.Figure 5 used normalized histograms and Gaussian fits for pixel-wise absolute-intensity errors.
6 Conclusion
The study demonstrates unsupervised brain-lesion detection by learning healthy MRI distributions and detecting abnormalities through reconstruction differences. A latent-consistency constraint improved the accuracy of abnormal-region detection, while AnoGAN comparison remained future work.
- Unsupervised detection: The approach learns prior knowledge from healthy brain data and detects abnormalities according to the learned healthy distribution.VAE and AAE models were investigated for unsupervised lesion detection.
- Latent constraint: A latent constraint was proposed to enforce consistency in latent representations and improve detection of abnormal regions.The paper reports improved detection accuracy with the proposed constraint.
- Scope of evaluation: Comparison with AnoGAN was deferred because the authors could not obtain decent results due to gradient-descent problems in latent space.The comparison was reserved for future work.