Unsupervised Lesion Detection via Image Restoration with a Normative Prior (2005.00031v1)

Abstract: Unsupervised lesion detection is a challenging problem that requires accurately estimating normative distributions of healthy anatomy and detecting lesions as outliers without training examples. Recently, this problem has received increased attention from the research community following the advances in unsupervised learning with deep learning. Such advances allow the estimation of high-dimensional distributions, such as normative distributions, with higher accuracy than previous methods.The main approach of the recently proposed methods is to learn a latent-variable model parameterized with networks to approximate the normative distribution using example images showing healthy anatomy, perform prior-projection, i.e. reconstruct the image with lesions using the latent-variable model, and determine lesions based on the differences between the reconstructed and original images. While being promising, the prior-projection step often leads to a large number of false positives. In this work, we approach unsupervised lesion detection as an image restoration problem and propose a probabilistic model that uses a network-based prior as the normative distribution and detect lesions pixel-wise using MAP estimation. The probabilistic model punishes large deviations between restored and original images, reducing false positives in pixel-wise detections. Experiments with gliomas and stroke lesions in brain MRI using publicly available datasets show that the proposed approach outperforms the state-of-the-art unsupervised methods by a substantial margin, +0.13 (AUC), for both glioma and stroke detection. Extensive model analysis confirms the effectiveness of MAP-based image restoration.

• The paper proposes a novel unsupervised method for brain lesion detection in MRI by formulating it as an image restoration problem using a normative prior.
• The method trains a latent variable model (like VAE or GMVAE) on healthy brain anatomy data and detects lesions as deviations from the model's reconstruction of the input image.
• Evaluations show the proposed framework significantly outperforms existing unsupervised techniques on glioma and stroke detection tasks, offering a scalable solution without requiring annotated lesion data.

Unsupervised Lesion Detection via Image Restoration with a Normative Prior

The paper in discussion proposes a novel approach to unsupervised lesion detection within medical imaging, specifically targeting brain MRI scans. A prevalent challenge in contemporary radiology is the accurate and automatic detection of brain lesions, which constitutes an essential step in diagnostic pipelines. Conventionally, lesion detection is performed manually, demanding substantial expertise and thorough understanding of brain anatomy. Despite the availability of supervised machine learning methods that excel in identifying predefined lesion types, these methods lack generalizability to unseen lesion types and require extensive annotated datasets for training. Thus, the focus of this paper is a framework for unsupervised detection that potentially offers a more scalable solution by removing the reliance on labeled data.

The authors introduce a novel probabilistic model that formulates the unsupervised lesion detection task as an image restoration problem. The central idea is to employ a network-based prior, capturing the normative distribution of healthy brain anatomy, to reconstruct the input image and identify lesions as deviations from the reconstructed images. The reconstruction is performed using Maximum-a-Posteriori (MAP) estimation, which aims to minimize false positives that typically emerge in such inverse problems by enforcing a regularization that penalizes strong deviations between the restored and the original images.

An extensive evaluation is conducted using public datasets containing gliomas and stroke lesions in brain MRI. The proposed model's efficacy is validated through a series of comparative experiments, demonstrating substantial improvements over existing unsupervised methods. The key metric used is the area under the receiver operating characteristic curve (AUC), with the newly proposed method outperforming state-of-the-art techniques by a margin of 0.13 for both glioma and stroke detection tasks.

The foundational principle underlying the suggested approach is the restoration framework, where the lesion detection process is cast into a probabilistic image restoration problem. The latent-variable model, either a Variational Autoencoder (VAE) or a Gaussian Mixture VAE (GMVAE), learns the distribution of healthy anatomy using T1 and T2 weighted MRI images of healthy subjects from the CamCAN dataset. The model's effectiveness hinges on its ability to fit complex high-dimensional distributions, aided by recent advances in deep learning.

During detection, the model reconstructs the image, projecting lesions as anomalies by contrasting the input and restored images. Importantly, the unsupervised nature stems from the model's training on healthy samples without any prior information on lesion-specific characteristics. Robustness is achieved by the TV-norm-based likelihood term in the MAP estimation, discouraging false lesion detections stemming from anatomical variations often inherent in patient images.

The strength of this framework lies not only in the impressive quantitative improvements but also in flexible adaptability to different types of lesions and neural network architectures. Future explorations in this domain might investigate model generalization across varying imaging modalities, refining the unsupervised learning paradigm for broader applications.

In summary, this research presents significant advancements in unsupervised lesion detection, promising enhanced diagnostic reliability and efficiency, particularly beneficial in clinical scenarios where lesion types are undefined or when comprehensive annotated training sets are unavailable. Continued research along these lines could catalyze substantial developments in the application of unsupervised learning for medical imaging, driving future innovations in diagnostic radiology.