Push-Forward Signed Distance Functions enable interpretable and robust continuous shape quantification

Roua Rouatbi, Juan Esteban Suarez, Ivo F. SbalzariniarXiv 2024 · arXiv · OpenReview

TL;DR

This paper introduces the Push-Forward Signed Distance Morphometric (PF-SDM), a new method for quantifying biological shapes that is continuous, robust to transformations, and interpretable. It outperforms existing methods in shape analysis and can even predict developmental outcomes in biological systems, offering a more powerful tool for understanding complex biological forms.

What Problem Does This Paper Address?

Understanding and comparing shapes is fundamental in many scientific fields, especially in biomedical imaging where subtle shape differences can reveal crucial biological information. However, existing methods for quantifying shape similarity often struggle with a few key challenges:

  • Invariance: Shapes can look the same even after transformations like rotation, translation, or scaling. Traditional methods often fail to be invariant to these transformations, leading to inaccurate comparisons.
  • Interpretability: Some methods produce complex, black-box representations that are hard to understand, making it difficult to glean biological insights.
  • Continuity and Smoothness: Many methods rely on discrete representations or approximations, which can limit their mathematical tractability and the ability to derive smooth gradients needed for advanced analysis.
  • Capturing Topology and Geometry: Effectively capturing both the overall form (geometry) and structural properties like holes or connections (topology) simultaneously remains a challenge.

This paper aims to bridge this gap by proposing a novel morphometric that is continuous, invariant to shape-preserving transformations, and provides interpretable features for robust shape analysis.

Methods

The PF-SDM is built upon the Signed Distance Function (SDF), a popular way to represent shapes by encoding the distance from any point in space to the shape's boundary. The authors introduce a novel "Push-Forward" step to achieve invariance. Here's a breakdown of their approach:

  1. Signed Distance Function (SDF) Computation: They start by computing a smooth SDF for a given closed shape. This is done by solving a viscous Eikonal equation, which ensures numerical stability and results in a smooth polynomial approximation of the SDF.
  2. Push-Forward Signed Distance Function (PF-SDF): This is the core innovation. To make the SDF invariant to transformations, they "push forward" the SDF onto a common reference domain (a unit disk). This is achieved by defining a deformation map that aligns the shape's boundary with the reference shape's boundary. This process effectively "warps" the SDF so that it's no longer dependent on the original position or orientation of the shape.
  3. Morphometric Evaluation: From the PF-SDF, they derive the PF-SDM by computing its Fourier transform. This transforms the shape representation into a set of coefficients that are invariant to rotation and reflection. These coefficients form a vector embedding of the shape, making it suitable for machine learning tasks. The PF-SDM can also optionally incorporate intensity information (like gene expression) from the image, creating an "intensity-aware" morphometric.

Datasets

  • Synthetic Data: A synthetic benchmark dataset was created by taking five base shapes (circle, elongated, folded, rounded square, branched) and applying ten random shape-preserving transformations (translation, rotation, reflection, scaling) to each. This resulted in 50 transformed shapes for evaluating robustness.
  • Mouse Gastruloids: Real biological data from time-lapse microscopy videos of mouse gastruloids was used. This dataset included 78 gastruloids, categorized by whether they developed a single axis or multiple axes. Images were captured in brightfield and with an mCherry-labeled Brachyury reporter.

Key Insights and Ideas

  • Robustness to Transformations: The PF-SDM demonstrably groups transformed instances of the same shape more tightly than traditional methods like EFA and GPA. This is a crucial step towards reliable shape comparison.
  • Interpretable Hierarchy: The PF-SDM reveals a meaningful hierarchy of shape similarity that mirrors typical developmental progressions in biological systems (e.g., from spherical to branched shapes). This suggests the morphometric captures biologically relevant geometric features.
  • Predictive Power in Biology: The PF-SDM was successfully applied to predict body-axis formation in mouse gastruloids, outperforming a CNN baseline. This highlights its utility in complex biological scenarios where subtle shape and intensity cues are important.
  • Fusion of Shape and Intensity: The ability to fuse shape features with intensity information (like gene expression) provides a more comprehensive understanding of biological processes, as seen in the gastruloid example.

Key Results

  • Synthetic Benchmark: PF-SDM showed superior robustness in clustering transformed shapes compared to EFA and GPA, as visualized by Multidimensional Scaling (MDS).
  • Mouse Gastruloids Classification: Logistic regression using PF-SDM features achieved higher accuracy and balanced accuracy than a CNN baseline for predicting gastruloid axis formation.
  • Computational Efficiency: PF-SDM-based classification required significantly less computation time (minutes vs. hours for CNN training).
  • Accuracy Gains: The accuracy improvements in the gastruloid task were primarily attributed to the shape features derived from PF-SDM, while intensity features performed comparably to the CNN.

Limitations and Future Work

  • Dimensionality: The current work focuses on 2D shapes and their temporal dynamics. Extending the PF-SDF framework to 3D+time is a natural next step.
  • Manifold Complexity: Generalizing from closed shapes to open or non-singly-connected manifolds (which might require vector SDFs) is noted as a more challenging area.
  • Higher-Order Features: Exploring the use of Sobolev norms could potentially capture higher-order geometric features.
  • Formal Proofs: The authors suggest formally proving the differentiability of PF-SDM distances, which could open new avenues for incorporating PF-SDM terms into machine learning loss functions.

Summary updated

Abstract

We introduce the Push-Forward Signed Distance Morphometric (PF-SDM), a novel method for shape quantification in biomedical imaging that is continuous, interpretable, and invariant to shape-preserving transformations. PF-SDM effectively captures the geometric properties of shapes, including their topological skeletons and radial symmetries. This results in a robust and interpretable shape descriptor that generalizes to capture temporal shape dynamics. Importantly, PF-SDM avoids certain issues of previous geometric morphometrics, like Elliptical Fourier Analysis and Generalized Procrustes Analysis, such as coefficient correlations and landmark choices. We present the PF-SDM theory, provide a practically computable algorithm, and benchmark it on synthetic data.
BibTeX
@misc{rouatbi2024push,
  author = {Roua Rouatbi and Juan Esteban Suarez and Ivo F. Sbalzarini},
  title = {{Push-Forward Signed Distance Functions enable interpretable and robust continuous shape quantification}},
  year = {2024},
  eprint = {2410.21004},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2410.21004},
}
Figure