Deep Autoencoding Models for Unsupervised Anomaly Segmentation in Brain MR Images (1804.04488v1)

Abstract: 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.

• The paper introduces AnoVAEGAN, which combines spatial VAEs with adversarial training to segment anomalies via pixel-wise reconstruction error.
• The method models the distribution of healthy brains to detect deviations, overcoming the limitations of traditional patch-based approaches.
• Experiments show that this unsupervised technique outperforms classical autoencoders, yielding higher Dice-scores in MS lesion segmentation.

Unsupervised Anomaly Segmentation in Brain MR Images with Deep Autoencoding Models

The paper "Deep Autoencoding Models for Unsupervised Anomaly Segmentation in Brain MR Images" by Christoph Baur et al. presents an innovative approach for the detection and segmentation of pathologies such as Multiple Sclerosis (MS) lesions in brain Magnetic Resonance (MR) images. The authors propose leveraging deep autoencoding models, particularly focusing on spatial Variational Autoencoders (VAEs) combined with adversarial networks, which they refer to as AnoVAEGAN, as a method for unsupervised anomaly detection and segmentation.

Methodological Contributions

This research deviates from traditional supervised segmentation approaches that require extensive labeled datasets, which are often challenging to source and may not encompass the full spectrum of pathologies. The paper posits that instead of relying on such datasets, modeling the distribution of healthy brains through unsupervised learning allows for the identification of anomalies as deviations from recognized normal patterns.

Key contributions of this work include:

1. Spatial Autoencoding: Unlike patch-based models, which operate on local image patches, the presented approach operates directly on entire 2D MR slices. Spatial VAEs are employed to capture "global" anatomical features of the brain, thereby avoiding the limitations of patch-based models.
2. Generative Adversarial Training: The authors combine VAEs with adversarial networks, aiming to enhance the realism and quality of reconstructions. This adversarial training component helps mitigate typical autoencoder pitfalls such as blurriness and memorization, with the decoder being trained to generate images capable of deceiving a discriminator network.
3. Anomaly Detection via Reconstruction Error: The method proposes computing the pixel-wise reconstruction error between the input image and its reconstruction to detect and delineate anomalies. This approach relies on the premise that the model will only reconstruct healthy structures accurately, leaving lesions poorly reconstructed, detectable by high error values.

Results and Implications

Experiments were conducted on an in-house dataset comprising normal and MS-lesional brain images. The AnoVAEGAN outperformed other compared methods, including classical Autoencoders (dAE), spatial Autoencoders (sAE), and the AnoGAN framework, in terms of Dice-score, a metric used to gauge the accuracy of medical image segmentations. Importantly, spatial models significantly outperformed traditional dense models, emphasizing the importance of preserving spatial information throughout the encoding and decoding process.

Practically, this research provides a compelling case for clinicians seeking to identify abnormalities without relying on extensive labeled datasets. The model's capability to operate quickly and effectively makes it a valuable tool in environments where swift diagnosis is crucial. Theoretically, this paper champions the integration of adversarial strategies within unsupervised frameworks, suggesting potential for expanded use in other medical imaging contexts.

Speculations on Future Developments

Future research could focus on extending these models to 3D representations to naturally capture volumetric data rather than relying on 2D slices, potentially improving detection accuracy. Additionally, leveraging richer statistical models such as Gaussian Mixture Models for the latent space could enhance the model's ability to capture the diversity of normal anatomical structures, further refining the distinction of anomalies.

Overall, this paper contributes significantly to the advancement of unsupervised anomaly segmentation in medical imaging, showcasing the effectiveness of integrating deep generative models with adversarial training to push the boundaries of automatic pathology detection.