Schroedinger Eigenmaps for the Analysis of Biomedical Data

Wojciech Czaja, Martin EhlerIEEE TPAMI 2013 · DOI

Introduction

Modern biomedical research generates high-dimensional datasets, such as multispectral retinal scans and gene expression profiles, which are often characterized by low signal-to-noise ratios. Unsupervised manifold learning techniques, like Laplacian Eigenmaps (LE), are adept at identifying underlying data structures but often fail to distinguish between overlapping classes. Conversely, supervised methods like Support Vector Machines (SVM) require substantial labeled data, which is costly and time-consuming to obtain in clinical settings.

To address this, the authors propose Schroedinger Eigenmaps (SE), a semi-supervised learning framework that incorporates sparse expert knowledge to "steer" dimensionality reduction. By replacing the standard graph Laplacian with a Schroedinger operator, the method effectively separates complex data even when labeled samples constitute less than 0.1% of the dataset.

Methodology: Steering the Manifold

The core of the SE algorithm is the integration of a "barrier potential" into the manifold learning process.

From Laplacian to Schroedinger

In traditional LE, a graph is constructed to represent local data similarities, and the eigenvectors of the graph Laplacian ($L$) provide a low-dimensional embedding. SE generalizes this by using a Schroedinger operator: $E = L + \alpha V$. Here, $V$ is a potential matrix representing expert labels, and $\alpha$ is a weight parameter controlling the influence of these labels.

The evolution of a manifold embedding under varying barrier potentials, illustrating the "steering" effect of Schroedinger Eigenmaps.

Barrier Potentials

The potential matrix $V$ can be configured in two primary ways:

  • Diagonal Potentials: High values are assigned to specific labeled points, "pushing" them toward the origin in the embedded space. This creates a barrier that helps isolate pathological or target classes from the rest of the data.
  • Non-diagonal Potentials: These can be used to "glue" disparate regions of a class together, forcing them to be close in the embedding even if they are not neighbors in the original high-dimensional space.

Experimental Results and Applications

The authors demonstrated the efficacy of SE across several domains, most notably in clinical imaging and genomics.

Retinal Multispectral Imaging

A significant application of SE is the detection of drusen—yellow deposits under the retina that serve as early markers for age-related macular degeneration (AMD). In a dataset of 160,000 pixels, labeling just 50 points (0.04%) allowed SE to generate a highly accurate map of the pathology.

Detection and segmentation of drusen in retinal multispectral imaging using Schroedinger Eigenmaps.

SE successfully identified and sub-classified drusen into "center" and "outer" regions, a task where SVMs struggled due to the non-linear mixing of spectral signatures from different retinal layers.

Performance on Sparse Data

On the Cleveland Heart Disease dataset, SE achieved a 15% error rate with only 40 training samples, significantly outperforming LE (42%) and both Linear and Gaussian SVMs (19-21%). This highlights SE's robustness in scenarios with extremely limited ground truth.

Conclusion and Future Directions

Schroedinger Eigenmaps offer a powerful, flexible approach to semi-supervised manifold learning. By modifying the operator itself rather than just warping the distance metric, SE provides a more nuanced way to incorporate expert knowledge into data analysis. While the method shows great promise for biomedical data, future work is needed to automate the selection of parameters like $\alpha$ and to explore more complex potential configurations for highly variable data classes.

Summary updated

BibTeX
@article{czaja2013schroedinger,
  author = {Wojciech Czaja and Martin Ehler},
  title = {{Schroedinger Eigenmaps for the Analysis of Biomedical Data}},
  journal = {IEEE Trans. Pattern Anal. Mach. Intell.},
  year = {2013},
  doi = {10.1109/tpami.2012.270},
}
Figure