Source-linked AI summary
Deep Autoencoding Models for Unsupervised Anomaly Segmentation in Brain MR Images
Christoph Baur, Benedikt Wiestler, Shadi Albarqouni, Nassir Navab
TL;DR
The paper addresses unsupervised brain-lesion detection when labeled pathological data are limited and patch-based methods do not model whole-slice anatomy. It uses spatial autoencoding with reconstruction-error segmentation, finding that spatial models outperform dense bottlenecks and AnoGAN, with marginal gains from latent normality constraints and adversarial training.
Problem
Brain-lesion segmentation needs methods that can model healthy anatomy without vast labeled datasets and without relying only on local image patches.
Method
AnoVAEGAN uses a fully convolutional spatial VAEGAN to model whole brain MR slices and detect lesions from pixelwise reconstruction residuals.
Results
Spatial autoencoding models significantly outperform dense-bottleneck models and AnoGAN, while latent-space constraints and adversarial training provide marginal improvements over non-generative spatial models.
Takeaways & Limitations
Whole-slice spatial autoencoding can support fast unsupervised brain-lesion segmentation and may provide prior information for supervised deep learning.
Abstract
from arXiv · showhide
Reliably modeling normality and differentiating abnormal appearances from normal cases is a very appealing approach for detecting pathologies in medical images. A plethora of such unsupervised anomaly detection approaches has been made in the medical domain, based on statistical methods, content-based retrieval, clustering and recently also deep learning. Previous approaches towards deep unsupervised anomaly detection model patches of normal anatomy with variants of Autoencoders or GANs, and detect anomalies either as outliers in the learned feature space or from large reconstruction errors. In contrast to these patch-based approaches, we show that deep spatial autoencoding models can be efficiently used to capture normal anatomical variability of entire 2D brain MR images. A variety of experiments on real MR data containing MS lesions corroborates our hypothesis that we can detect and even delineate anomalies in brain MR images by simply comparing input images to their reconstruction. Results show that constraints on the latent space and adversarial training can further improve the segmentation performance over standard deep representation learning.
1 Introduction
The paper frames brain-lesion detection as unsupervised anomaly detection by modeling healthy anatomy, avoiding dependence on extensive pathological labels. It proposes spatial autoencoding of whole brain MR slices and identifies lesions through reconstruction differences.
- Supervised pathology segmentation requires vast labeled datasets and may generalize poorly because training data rarely covers all pathological appearances.
- The proposed alternative models the distribution of healthy brains and treats pathological appearances as deviations from that norm.
- Detection and delineation are obtained from pixel-wise reconstruction error, including by subtracting the reconstructed image from the input.
- Unlike prior patch-based approaches, the method performs deep convolutional representation learning on entire MR slices.
- AnoVAEGAN uses spatial VAEs to capture global normal anatomy and adversarially trains the decoder to improve reconstructed-sample realism.
- Experiments compare spatial and dense autoencoding variants with AnoGAN for unsupervised MS-lesion delineation, reporting significant improvements from spatial autoencoding models.
2 Methodology
The method encodes whole brain MR images into a structured spatial latent representation and reconstructs them with a VAEGAN-style encoder-decoder. Lesions are delineated from thresholded pixel-wise reconstruction residuals.
- The encoder maps an input image x to a lower-dimensional latent representation z, and the decoder maps z back to a high-quality reconstruction.
- The latent representation is constrained toward a multivariate normal distribution, while a discriminator distinguishes real images from reconstructed images.
- Training.: Training alternates optimization of the VAE component and discriminator using reconstruction, prior-regularization, and adversarial objectives.
- Training.: The reconstruction objective uses pixelwise ℓ1 distance, and the adversarial loss encourages the decoder to generate images that fool the discriminator.
- Training.: A fully convolutional encoder-decoder preserves spatial information by representing z as a multidimensional tensor.
- Anomaly Detection: Anomalies are segmented by computing pixelwise ℓ1 distance between input and reconstruction, then thresholding the residual image.
3 Experiments and Results
The experiments compare dense and spatial autoencoding models, latent-space configurations, and adversarial variants for unsupervised MS lesion segmentation. Evaluation uses Dice scores and reconstruction time, with residual postprocessing and visual comparisons across models.
- Experimental design: The study tests whether autoencoding networks can reconstruct complex brain MR images and examines latent dimensionality, latent-space structure, and adversarial training.These experiments compare reconstruction and segmentation behavior across several autoencoder variants.
- Dataset: The dataset contains FLAIR and T1 image pairs from 83 healthy patients and 49 patients with MS lesions, all rigidly co-registered to the SRI24 atlas.The healthy data provide anatomical variability for modeling normality, while lesion cases support evaluation.
- Evaluation: Performance is measured by mean and standard deviation of patient-level Dice scores and average time required to reconstruct and segment a sample.The evaluation therefore covers both segmentation quality and computational cost.
- Model comparison: Spatial autoencoders use a fully convolutional bottleneck that preserves spatial information, with experiments varying feature-map dimensionality and identifying z = 16 × 16×64 as a useful configuration.Dense bottlenecks and spatial bottlenecks are compared while keeping encoder-decoder complexity constant across architectures.
- Anomaly detection: Residual-based segmentations are postprocessed with median filtering, eroded brain masks, model-specific thresholding, and removal of connected components smaller than 6 voxels.These steps target small residuals and unlikely lesion candidates before producing the binary segmentation mask.
- Results: AnoVAEGAN achieves the highest average Dice score, while spatial autoencoding models generally outperform dense-bottleneck models and most AnoGAN variants.The spatial AE with adversarial training performs worse than its generative counterpart and other spatial models with the same bottleneck resolution.
4 Discussion and Conclusion
The paper concludes that generative deep representation learning can provide a fast unsupervised anomaly-detection approach for full brain MR slices. Its experiments support spatial autoencoding with sufficient bottleneck resolution, while future work considers richer latent distributions and 3D models.
- Discussion: Dense-bottleneck AE and VAE models fail to reconstruct anomalies while also losing fine brain-MR details, whereas spatial models with 16×16px feature maps mitigate this problem.An 8 × 8px bottleneck instead causes severe information loss and large reconstruction errors in general.
- Discussion: Constraining the latent space to a multivariate normal distribution and adding adversarial training yields marginal improvements over non-generative models.Spatial autoencoding also outperforms AnoGAN and reconstructs samples considerably faster.
- Conclusion: The proposed approach is a novel, fast unsupervised anomaly-detection method that encodes the full context of brain MR slices.The authors suggest it may also provide prior information for supervised deep learning.
- Future work: Future work will investigate Gaussian-mixture latent spaces and 3D autoencoding models for unsupervised anomaly detection.These directions extend the current single-Gaussian latent-space and 2D modeling choices.