Surreal-GAN:Semi-Supervised Representation Learning via GAN for uncovering heterogeneous disease-related imaging patterns

Zhijian Yang,Junhao Wen,Christos Davatzikos

DOI: https://doi.org/10.48550/arXiv.2205.04523

2022-05-10

Abstract:A plethora of machine learning methods have been applied to imaging data, enabling the construction of clinically relevant imaging signatures of neurological and neuropsychiatric diseases. Oftentimes, such methods don't explicitly model the heterogeneity of disease effects, or approach it via nonlinear models that are not interpretable. Moreover, unsupervised methods may parse heterogeneity that is driven by nuisance confounding factors that affect brain structure or function, rather than heterogeneity relevant to a pathology of interest. On the other hand, semi-supervised clustering methods seek to derive a dichotomous subtype membership, ignoring the truth that disease heterogeneity spatially and temporally extends along a continuum. To address the aforementioned limitations, herein, we propose a novel method, termed Surreal-GAN (Semi-SUpeRvised ReprEsentAtion Learning via GAN). Using cross-sectional imaging data, Surreal-GAN dissects underlying disease-related heterogeneity under the principle of semi-supervised clustering (cluster mappings from normal control to patient), proposes a continuously dimensional representation, and infers the disease severity of patients at individual level along each dimension. The model first learns a transformation function from normal control (CN) domain to the patient (PT) domain with latent variables controlling transformation directions. An inverse mapping function together with regularization on function continuity, pattern orthogonality and monotonicity was also imposed to make sure that the transformation function captures necessarily meaningful imaging patterns with clinical significance. We first validated the model through extensive semi-synthetic experiments, and then demonstrate its potential in capturing biologically plausible imaging patterns in Alzheimer's disease (AD).

Machine Learning,Computer Vision and Pattern Recognition,Image and Video Processing

What problem does this paper attempt to address?

The problem that this paper attempts to solve is that in neuroimaging, existing machine - learning methods are insufficient in dealing with disease heterogeneity. Specifically:

1. **Explicit Modeling of Disease Heterogeneity**: Many existing methods do not explicitly model disease heterogeneity or handle it through nonlinear models, but these models are often difficult to interpret.

2. **Limitations of Unsupervised Methods**: Unsupervised methods may resolve heterogeneity caused by confounding factors rather than heterogeneity related to specific pathologies.

3. **Limitations of Semi - supervised Clustering Methods**: Existing semi - supervised clustering methods usually attempt to divide patients into two subtypes, ignoring that disease heterogeneity is a continuous process in space and time.

To overcome the above problems, the authors propose a new method - Surreal - GAN (Semi - Supervised Representation Learning via GAN). The main goals of Surreal - GAN are:

- **Resolving Disease - related Heterogeneity**: Resolve disease - related heterogeneity through semi - supervised clustering from the normal control group (CN) to the patient group (PT).
- **Continuous - dimensional Representation**: Propose a continuous - dimensional representation method, where each dimension represents a relatively homogeneous imaging pattern and its severity.
- **Inference of Disease Severity at the Individual Level**: Infer the disease severity of patients at the individual level.

Surreal - GAN achieves these goals through the following key steps:

1. **Learning of Transformation Functions**: Learn a transformation function from the CN domain to the PT domain, where the latent variable controls the transformation direction.
2. **Inverse Mapping Function**: Introduce an inverse mapping function to ensure that the transformation function captures clinically meaningful imaging patterns.
3. **Regularization**: Guide the model to capture meaningful imaging patterns through various regularization methods (such as sparse transformation, Lipschitz continuity, pattern orthogonality, and monotonicity).

Through these methods, Surreal - GAN can capture biologically reasonable imaging patterns under different conditions and shows a significant correlation in experiments on Alzheimer's disease (AD).