Statistical multi-level shape models for scalable modeling of multi-organ anatomies

Nawazish Khan, Andrew C. Peterson, B. Aubert, Alan Morris, Penny R. Atkins, Amy L. Lenz, Andrew E. Anderson, Shireen ElhabianFrontiers in Bioengineering and Biotechnology 2023 · DOI

TL;DR

Researchers have developed a "Multi-Level Multi-Organ" (MLMO) framework that disentangles the shape of individual organs from their relative positions (pose). By splitting shape statistics into two independent mathematical subspaces, this approach creates more accurate, computationally efficient, and anatomically realistic models of complex structures like the spine and foot.

What Problem Does This Paper Address?

In medical imaging, Statistical Shape Modeling (SSM) is used to understand how anatomy varies across a population. A popular method is Particle-based Shape Modeling (PSM), which places "correspondence points" (landmarks) across a group of 3D organ models to learn their common patterns.

However, when dealing with multiple organs—like the 17 vertebrae in a human spine—traditional PSM treats the entire assembly as a single, massive object. This "global" approach has three major flaws:

  1. Entanglement: It mixes up changes in an organ's actual shape (morphology) with changes in how organs are arranged (pose). This makes it hard for doctors to distinguish between a deformed bone and a misaligned joint.
  2. Anatomical Errors: Because the model doesn't "know" the organs are separate, it often produces unrealistic results, such as bones overlapping or jagged, "broken" reconstructions.
  3. Scalability: As you add more organs, the math becomes exponentially more complex, leading to a massive computational "memory footprint."

Methods

The authors propose the Multi-Level Multi-Organ (MLMO) model. The core innovation is leveraging Multilevel Component Analysis (MLCA) to split the shape space into two mutually orthogonal (independent) subspaces:

  • Within-organ subspace: Captures the morphological "look" of each individual bone.
  • Between-organ subspace: Captures the "pose" or relative alignment between the bones.

Instead of one giant optimization, the algorithm alternates between these two spaces. This disentanglement allows the model to optimize landmarks on a vertebra-by-vertebra basis while simultaneously maintaining the integrity of the entire spinal curve. Mathematically, this replaces one massive correlation matrix with several much smaller ones, significantly speeding up the process and reducing the data needed to create a reliable model.

Datasets

The researchers validated their method using both synthetic and real-world clinical data:

  • Synthetic "Supershapes": Parameterized 3D objects used as a proof-of-concept to show the model could perfectly separate shape from pose.
  • Spinal Column: 30 subjects from the public VerSe dataset, focusing on 17 vertebrae (T1–L5).
  • Foot and Ankle: Weightbearing CT (WBCT) scans of 27 participants, modeling four key bones (talus, calcaneus, navicular, and cuboid).
  • Hip Joint: 51 pairs of hemi-pelvis and proximal femurs, including healthy controls and patients with impingement or dysplasia.

Key Insights and Ideas

  • Disentanglement equals Clinical Utility: By separating pose from shape, the model can automatically detect clinical markers like "pelvic tilt" or "joint space narrowing" without getting confused by the actual size or shape of the bones.
  • The "Aha" Moment on Scalability: Traditional models struggle with "high-dimensional, low-sample-size" problems (lots of landmarks, few patients). By breaking the problem into subspaces, MLMO becomes much more "flexible" and requires fewer subjects to reach statistical power.
  • Preserving Anatomy: Because the model treats organs as separate entities in the "within-organ" space, it eliminates the "jagged" artifacts and bone-clipping common in global models. The reconstructed bones look like real bones.

Key Results

The MLMO model outperformed the standard global approach across all standard metrics:

  • Compactness & Generalization: The MLMO model explained more anatomical variation with fewer "modes" (variables) and was significantly better at accurately reconstructing "unseen" anatomy not present in the training set.
  • Age Prediction: In a regression task using spinal data, the MLMO model predicted a patient's age with an $R^2$ of 0.62, compared to just 0.20 for the global model. This suggests the MLMO model captures the subtle morphological changes of aging (like vertebral endplate growth) much more effectively.
  • Joint Coverage: In the foot and ankle, the MLMO model was more accurate at predicting the "coverage area" of the subtalar joint, a critical metric for diagnosing instability and osteoarthritis.

Limitations and Future Work

The authors acknowledge that their current model assumes a linear relationship (Gaussian distribution) for pose variations. In reality, joint movements (like the swinging of a hip) are often non-linear or rotational.

Future work aims to:

  • Incorporate non-linear embeddings to better model complex joint articulations.
  • Develop "conditional" models that can predict the shape of a missing organ based on the known parts of the multi-organ complex.
  • Automate the placement of clinical landmarks for surgical planning based on these disentangled subspaces.

Summary updated

Abstract

Statistical shape modeling is an indispensable tool in the quantitative analysis of anatomies. Particle-based shape modeling (PSM) is a state-of-the-art approach that enables the learning of population-level shape representation from medical imaging data (e.g., CT, MRI) and the associated 3D models of anatomy generated from them. PSM optimizes the placement of a dense set of landmarks (i.e., correspondence points) on a given shape cohort. PSM supports multi-organ modeling as a particular case of the conventional single-organ framework <i>via</i> a global statistical model, where multi-structure anatomy is considered as a single structure. However, global multi-organ models are not scalable for many organs, induce anatomical inconsistencies, and result in entangled shape statistics where modes of shape variation reflect both within- and between-organ variations. Hence, there is a need for an efficient modeling approach that can capture the inter-organ relations (i.e., pose variations) of the complex anatomy while simultaneously optimizing the morphological changes of each organ and capturing the population-level statistics. This paper leverages the PSM approach and proposes a new approach for correspondence-point optimization of multiple organs that overcomes these limitations. The central idea of multilevel component analysis, is that the shape statistics consists of two mutually orthogonal subspaces: the within-organ subspace and the between-organ subspace. We formulate the correspondence optimization objective using this generative model. We evaluate the proposed method using synthetic shape data and clinical data for articulated joint structures of the spine, foot and ankle, and hip joint.
BibTeX
@inproceedings{khan2023statistical,
  author = {Nawazish Khan and Andrew C. Peterson and B. Aubert and Alan Morris and Penny R. Atkins and Amy L. Lenz and Andrew E. Anderson and Shireen Elhabian},
  title = {{Statistical multi-level shape models for scalable modeling of multi-organ anatomies}},
  booktitle = {Frontiers in Bioengineering and Biotechnology},
  year = {2023},
  doi = {10.3389/fbioe.2023.1089113},
}
Figure