IMODAL: Creating Learnable User-Defined Deformation Models

Leander Lacroix, Benjamin Charlier, Alain Trouvé, Barbara GrisCVPR 2021 · DOI

Abstract

Shape registration is a fundamental task in computational anatomy, traditionally dominated by generic models like Large Deformation Diffeomorphic Metric Mapping (LDDMM). However, these models often lack the structural constraints necessary to represent specific biological processes accurately. We present IMODAL (Implicit Modular Object Deformation Analysis Library), a framework that allows users to define "deformation modules" to bake prior knowledge directly into registration. This approach yields physically plausible and interpretable results, as demonstrated in biological growth modeling and image disentanglement tasks.

Introduction: The Need for Structured Registration

In the field of computational anatomy, registration is the process of finding a mathematical transformation that maps a source shape onto a target. For years, the gold standard has been LDDMM, which provides smooth, reversible warps. However, LDDMM is "all-purpose" and "weakly structured." When modeling biological processes like leaf growth, LDDMM may produce mathematically valid transformations that fail to reflect actual growth patterns, such as expansion from the base (basipetal) or the tip (acropetal). It lacks a specific "vocabulary" for movements like local stretching or shearing, and its high-dimensional outputs are often difficult for researchers to interpret. IMODAL addresses this gap by allowing users to define the rules of deformation.

Methods: The IMODAL Framework

The authors introduce IMODAL (Implicit Modular Object Deformation Analysis Library), a Python-based framework built on the concept of Deformation Modules. This modular approach allows users to define the rules of deformation rather than just the final state.

Modular Architecture

Users can combine multiple modules to describe complex transformations. For example, one module might handle global translation, while another handles local growth.

  • Implicit Modules: Instead of defining exact point movements, users define constraints. Order 0 Modules constrain vector field values (e.g., local translations), while Order 1 Modules constrain derivatives. This allows for modeling the "infinitesimal strain tensor," capturing how an object expands, dilates, or rotates along specific axes.
  • Silent Modules: These carry geometric information—such as the orientation of a leaf's veins—that evolves with the shape without generating its own force.
  • Hamiltonian Dynamics: Registration is framed as an energy minimization problem. By optimizing "initial momentum," the system finds the most efficient path between shapes that stays within the allowed "vocabulary" of the chosen modules.

The library is implemented using PyTorch and KeOps, enabling efficient GPU computation for large 2D and 3D datasets.

Datasets and Results

The framework was evaluated using several specialized datasets, most notably real-world leaf growth data.

Real-world leaf growth data showing (a) basipetal (growth from base) and (b) acropetal (growth from tip) patterns.

Biological Accuracy and Interpretability

In leaf growth tests, standard LDDMM failed to capture the basipetal nature of the growth, often resulting in generic warping that ignored biological reality. IMODAL, equipped with an Order 1 module, successfully recovered the growth pattern even without point-to-point correspondence.

Comparison of IMODAL and LDDMM registration over time (t=0 to t=1). IMODAL maintains biological plausibility by following the growth constraints.

Beyond biological accuracy, IMODAL provides superior interpretability. Because deformations are driven by a small number of user-defined "controls," the output is a set of curves showing how different parts of the object grew or shrunk over time. Furthermore, IMODAL can "learn" intrinsic local frames—such as the principal axes of growth—during the registration process, effectively uncovering the internal structure of the object from the deformation itself.

Conclusion

IMODAL bridges the gap between generic "black-box" warping and physically constrained modeling. By restricting the model's "vocabulary," the results become more meaningful and biologically relevant. While the framework requires users to select appropriate modules, it offers a powerful new tool for modeling complex developments in biological and medical imaging, such as tumor growth or organ development, where specific physical constraints are well-known but difficult to implement in traditional software.

Summary updated

BibTeX
@inproceedings{lacroix2021imodal,
  author = {Leander Lacroix and Benjamin Charlier and Alain Trouvé and Barbara Gris},
  title = {{IMODAL: Creating Learnable User-Defined Deformation Models}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2021},
  doi = {10.1109/cvpr46437.2021.01271},
}
Figure