Patient-Specific Conditional Joint Models of Shape, Image Features and Clinical Indicators

Bernhard Egger, Markus D. Schirmer, Florian Dubost, Marco J. Nardin, Natalia S. Rost, Polina Golland

TL;DR

Researchers have developed a new framework that jointly models 3D anatomical shapes, image features, and clinical data (like age or stroke scores) using Gaussian copulas. This approach allows clinicians to generate patient-specific anatomical models by conditioning on available clinical indicators, significantly improving our ability to predict disease progression and understand the interplay between biology and clinical outcomes.

What Problem Does This Paper Address?

In medical imaging, we often have two parallel streams of data: high-dimensional image data (like the 3D shape of a brain ventricle) and tabular clinical indicators (like a patient’s age, sex, or smoking history). While these are clearly related—for instance, ventricles often expand with age or disease—they are rarely analyzed together in a single statistical framework.

The technical challenge lies in the "data soup." Shape data consists of continuous 3D coordinates, while clinical data is a messy mix of binary (sex), ordinal (stroke severity scales), and continuous (age) variables. Most traditional statistical shape models assume everything follows a Gaussian (bell curve) distribution, which fails miserably when applied to discrete clinical categories. This paper addresses the need for a flexible, joint model that can handle these diverse data types simultaneously.

Methods

The authors employ a mathematical tool called a copula. Think of a copula as the "statistical glue" that separates the individual behavior of each variable (the marginal distributions) from the way they depend on each other (the joint structure).

  1. Marginal Modeling: First, they model each variable independently. For shape coordinates, they might use a Gaussian distribution; for clinical scores, they use empirical distributions based on the data.
  2. The Latent Space: They transform all these different variables into a shared "latent space" where every variable looks like a standard Gaussian. To handle non-continuous data (like a "yes/no" for hypertension), they use a clever ranking and sampling strategy to map these discrete values into the continuous latent space.
  3. Gaussian Processes (GP): Once everything is in this unified Gaussian space, they use Gaussian Process regression to capture the dependencies.
  4. Conditioning: Because the model is generative, you can "plug in" what you know (e.g., "This patient is an 80-year-old male") and the model will output a conditional distribution for the missing parts (e.g., "Here is the most likely shape of his lateral ventricles and the expected distribution of white matter disease").

Datasets

The study utilized a dataset of 793 acute ischemic stroke patients from the Genes Affecting Stroke Risk and Outcomes Study (GASROS).

  • Imaging: T2-FLAIR MRI scans were used to segment the lateral ventricles (represented as a mesh with 1,504 vertices) and White Matter Hyperintensities (WMH).
  • Clinical Indicators: Nine variables were included, such as age, sex, hypertension, smoking status, and stroke outcome scores (NIHSS and mRS).

Key Insights and Ideas

  • Decoupling is Key: By using copulas, the researchers aren't forced to assume that "smoking status" follows a bell curve. They can model the clinical reality of the data while still using the powerful machinery of Gaussian Processes to find correlations.
  • Shape as a Clinical Predictor: The model reveals that anatomical shape and the spatial distribution of white matter lesions are strong indicators of age, but interestingly, they contribute less to predicting immediate stroke outcome scores.
  • Patient-Specific Priors: Instead of comparing a patient to a "universal average," this model allows for the creation of a "patient-specific average." If you know a patient's demographic profile, the model provides a much more accurate baseline for what their brain should look like, making it easier to spot anomalies.

Key Results

The authors evaluated the model by seeing how well it could "fill in the blanks" (reconstruction tasks):

  • Clinical Prediction: Predicting a patient's age using only the "mean" resulted in an error of 15.77 years. By conditioning the model on ventricle shape and WMH burden, that error dropped to 9.85 years.
  • Shape Prediction: When predicting the 3D coordinates of the ventricles, including clinical indicators significantly reduced the error compared to using a standard unconditional mean.
  • Qualitative Impact: The researchers demonstrated that conditioning on "Age = 80" vs "Age = 40" produced visibly different mean shapes and WMH distributions, capturing the known biological reality of brain aging within a mathematical framework.

Limitations and Future Work

The authors acknowledge two primary hurdles:

  1. Correspondence: The model relies on "dense correspondence," meaning every 3D point on one patient's ventricle must be correctly mapped to the same point on another's. This is computationally expensive and difficult for highly variable pathologies.
  2. Linear Dependencies: The Gaussian copula primarily captures second-order dependencies (correlations). It may miss more complex, non-linear relationships between clinical factors and anatomy.

Future work could involve moving beyond the Gaussian copula to more complex dependency structures or applying this framework to other organs and diseases where clinical context is vital for image interpretation.

Summary updated

Abstract

We propose and demonstrate a joint model of anatomical shapes, image features and clinical indicators for statistical shape modeling and medical image analysis. The key idea is to employ a copula model to separate the joint dependency structure from the marginal distributions of variables of interest. This separation provides flexibility on the assumptions made during the modeling process. The proposed method can handle binary, discrete, ordinal and continuous variables. We demonstrate a simple and efficient way to include binary, discrete and ordinal variables into the modeling. We build Bayesian conditional models based on observed partial clinical indicators, features or shape based on Gaussian processes capturing the dependency structure. We apply the proposed method on a stroke dataset to jointly model the shape of the lateral ventricles, the spatial distribution of the white matter hyperintensity associated with periventricular white matter disease, and clinical indicators. The proposed method yields interpretable joint models for data exploration and patient-specific statistical shape models for medical image analysis.
BibTeX
@inproceedings{egger2019patient,
  author = {Bernhard Egger and Markus D. Schirmer and Florian Dubost and Marco J. Nardin and Natalia S. Rost and Polina Golland},
  title = {{Patient-Specific Conditional Joint Models of Shape, Image Features and Clinical Indicators}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2019},
  url = {https://openreview.net/forum?id=HkoIJBHgAM},
}

Loading related content...

Figure