Learning spatiotemporal statistical shape models for non-linear dynamic anatomies

Jadie Adams, Nawazish Khan, Alan Morris, Shireen ElhabianFrontiers in Bioengineering and Biotechnology 2023 · DOI

Abstract

Understanding anatomical shape changes over time is critical for clinical investigations, ranging from dynamic organ cycle characterization to longitudinal disease progression analysis. Spatiotemporal statistical shape modeling (SSM) provides a framework for evaluating these variations within a population. However, existing data-driven SSM approaches often assume sample independence or linear dynamics, making them unsuitable for complex, non-linear biological motions. This paper presents a principled approach to spatiotemporal SSM that incorporates underlying time dependency into correspondence optimization via regularized principal component polynomial regression. By operating within a reduced PCA subspace, the method captures non-linear temporal dynamics while maintaining spatial regularity. Validation on synthetic and cardiac MRI data demonstrates that this approach significantly outperforms existing methods in capturing population modes of variation and statistically significant time dependencies.

Introduction

In clinical research, the ability to quantify how an organ moves—such as the beating of a heart—or how a disease progresses over years is vital. Statistical Shape Models (SSMs) are used to convert complex 3D anatomical structures into compact mathematical representations. Traditional SSMs often suffer from "time-blindness," treating sequential scans of the same patient as independent data points. This cross-sectional approach ignores within-subject correlations. Furthermore, existing spatiotemporal models frequently assume linear shape changes, which fail to represent the cyclical and non-linear nature of biological processes like the cardiac cycle.

Methodology

The proposed method integrates Regularized Principal Component Polynomial Regression directly into the shape-learning process to address these limitations. The workflow consists of several key stages:

1. Point Distribution Models (PDMs)

The process begins with the establishment of "particles" or landmarks that automatically find corresponding locations across different subjects' scans. This is achieved through an entropy-based optimization scheme that ensures geometrically consistent particle placement across the cohort.

Example of a Point Distribution Model (PDM) with 128 particles across three ellipsoid shapes. Color denotes correspondence, showing consistent placement as the shape deforms.

2. PCA Subspace and Polynomial Regression

Rather than tracking individual $(x, y, z)$ coordinates for thousands of particles, which can lead to "jittery" models, the data is projected into a lower-dimensional "shape space" using Principal Component Analysis (PCA). Within this subspace, a polynomial function is fitted to describe shape evolution over time. This allows the model to follow the non-linear "curves" of motion, such as the expansion and contraction of a heart chamber, while ensuring that neighboring parts of the anatomy move together realistically.

3. Regularization and Optimization

To prevent overfitting to noise in medical imaging, the authors apply Elastic Net regularization. The algorithm uses an interleaved optimization strategy, alternating between moving particles to better represent the anatomy and updating the regression model to accurately capture the motion trajectory.

Experimental Validation

The method was validated using two distinct datasets:

  • Synthetic Ellipsoids: A cohort of 30 subjects with 8 time points each. The shapes were designed with subject-specific variation on one axis and sinusoidal temporal variation on another, testing the model's ability to disentangle these factors.
  • 4D Left Atrium: Real-world data from 28 patients with 25 time points covering a full cardiac cycle, segmented from CINE MRI scans. This dataset represents a significant challenge due to the highly non-linear filling and pumping phases.

Key Results and Discussion

The proposed approach demonstrated clear superiority over baseline cross-sectional and linear models:

  • Statistical Significance: In the left atrium experiment, the method was the only one to capture a statistically significant time dependency ($p < 0.001$). Baseline methods failed to reject the null hypothesis ($p > 0.05$), essentially failing to "see" the heart beating.
  • Accuracy: On synthetic data, the model achieved an $R^2$ value of 0.999, indicating near-perfect capture of the intended motion.
  • Spatial Regularity: By performing regression in the PCA subspace, the model maintained a "spatial prior," ensuring that the resulting modes of variation were biologically plausible and smooth rather than erratic.

Conclusion

By combining statistical shape modeling with non-linear polynomial regression in a reduced subspace, this research provides a robust framework for modeling dynamic anatomies. The ability to handle sparse data and non-linear trajectories makes it a powerful tool for longitudinal clinical studies. Future work will explore Mixed-Effects Modeling to better distinguish between average population trends and individual patient variations, as well as developing generative capabilities to predict missing time points in clinical sequences.

Summary updated

Abstract

Numerous clinical investigations require understanding changes in anatomical shape over time, such as in dynamic organ cycle characterization or longitudinal analyses (e.g., for disease progression). Spatiotemporal statistical shape modeling (SSM) allows for quantifying and evaluating dynamic shape variation with respect to a cohort or population of interest. Existing data-driven SSM approaches leverage information theory to capture population-level shape variations by learning correspondence-based (landmark) representations of shapes directly from data using entropy-based optimization schemes. These approaches assume sample independence and thus are unsuitable for sequential dynamic shape observations. Previous methods for adapting entropy-based SSM optimization schemes for the spatiotemporal case either utilize a cross-sectional design (ignoring within-subject correlation) or impose other limiting assumptions, such as the linearity of shape dynamics. Here, we present a principled approach to spatiotemporal SSM that relaxes these assumptions to correctly capture population-level shape variation over time. We propose to incorporate modeling the underlying time dependency into correspondence optimization <i>via</i> a regularized principal component polynomial regression. This approach is flexible enough to capture non-linear temporal dynamics while encoding population-specific spatial regularity. We demonstrate our method's efficacy on synthetic data and left atrium segmented from cardiac MRI scans. Our approach better captures the population modes of variation and a statistically significant time dependency than existing methods.
BibTeX
@inproceedings{adams2023learning,
  author = {Jadie Adams and Nawazish Khan and Alan Morris and Shireen Elhabian},
  title = {{Learning spatiotemporal statistical shape models for non-linear dynamic anatomies}},
  booktitle = {Frontiers in Bioengineering and Biotechnology},
  year = {2023},
  doi = {10.3389/fbioe.2023.1086234},
}
Figure