Quantification and visualization of variation in anatomical trees

Nina Amenta, Manasi Datar, Asger Dirksen, Marleen de Bruijne, Aasa Feragen, Xiaoyin Ge, Jesper Holst Pedersen, Marylesa Howard, Megan Owen, Jens Petersen, Jie Shi, Qiuping XuarXiv 2014 · DOI · arXiv

Abstract: This paper presents two approaches to quantifying and visualizing variation in datasets of trees. The first approach localizes subtrees in which significant population differences are found through hypothesis testing and sparse classifiers on subtree features. The second approach visualizes the global metric structure of datasets through low-distortion embedding into hyperbolic planes in the style of multidimensional scaling. A case study is made on a dataset of airway trees in relation to Chronic Obstructive Pulmonary Disease (COPD).

Introduction

Anatomical trees—such as lung airways, blood vessel networks, or neuronal dendrites—present a unique challenge for statistical analysis. Unlike data in standard Euclidean space, these structures vary simultaneously in topology (branching patterns) and geometry (branch lengths and shapes). They reside in a non-linear, negatively curved "tree-space" where traditional averaging methods fail. This paper introduces a robust framework to quantify population differences (e.g., healthy vs. diseased) and visualize global variation in these complex datasets.

Methodology

The authors propose two primary strategies for analyzing tree-structured data:

1. Local Subtree Analysis

Rather than analyzing the tree as a whole or as isolated branches, the framework focuses on nested subtrees. Permutation tests are employed to identify subtrees with statistically significant differences in mean shape or variance between groups. To ensure clinical interpretability, the researchers apply structured sparsity via $L_1$-regularized (Lasso) and Elastic Net logistic regression. This approach forces the model to select only the most predictive subtrees, pinpointing specific anatomical regions affected by disease.

Anatomical structure of the airway tree with highlighted subtrees for local analysis.

2. Global Visualization via Hyperbolic Embedding

Standard Multidimensional Scaling (MDS) often creates massive distortion when squashing high-dimensional tree-space data into a flat 2D Euclidean plane. Because tree-space has negative curvature, the authors utilize Hyperbolic MDS (HMDS) to embed data into a hyperbolic disc. This space "expands" exponentially toward its edges, naturally accommodating the geometry of branching trees and providing a more accurate representation of population relationships.

Datasets

  • The Airway Dataset: 600 airway trees extracted from pulmonary CT scans (300 COPD subjects and 300 healthy controls) as part of the Danish Lung Cancer Screening Trial.
  • Synthetic Benchmarks: Models named CORNER and SHEETS were designed to mimic mathematical singularities in tree-space, such as "open books" where multiple Euclidean regions meet along a common spine.

Key Results

The local analysis successfully localized COPD-related changes to specific regions, notably the Right Main Bronchus (RMB) and the Bronchus Intermedius (BrInt), which showed highly significant differences in mean shape.

P-values indicating significant differences in mean and variance across various airway subtrees.

Key findings include:

  • Classification Performance: Sparse logistic regression on subtree similarities achieved a classification accuracy of 65%, significantly outperforming standard classifiers like SVM or LDA (54-56%).
  • Subtrees vs. Branches: Analyzing subtrees proved far more effective than analyzing individual branches, suggesting that disease-related changes in the lungs are cumulative and structural.
  • Hyperbolic Advantage: Quantitative tests showed that Hyperbolic MDS consistently produced lower multiplicative distortion compared to Euclidean counterparts, particularly on complex synthetic datasets and the real COPD data.

Discussion and Future Work

While the 65% classification accuracy is a notable improvement, it remains below the threshold for standalone clinical diagnosis. A primary limitation is the high correlation between overlapping nested subtrees, which can complicate feature selection in Lasso models. Future work will focus on developing specialized "structured sparsity" methods that explicitly account for hierarchical parent-child relationships during feature selection, and exploring visualization spaces that mirror the stratified nature of tree-space even more closely.

Summary updated

Abstract

This paper presents two approaches to quantifying and visualizing variation in datasets of trees. The first approach localizes subtrees in which significant population differences are found through hypothesis testing and sparse classifiers on subtree features. The second approach visualizes the global metric structure of datasets through low-distortion embedding into hyperbolic planes in the style of multidimensional scaling. A case study is made on a dataset of airway trees in relation to Chronic Obstructive Pulmonary Disease.
BibTeX
@misc{amenta2014quantification,
  author = {Nina Amenta and Manasi Datar and Asger Dirksen and Marleen de Bruijne and Aasa Feragen and Xiaoyin Ge and Jesper Holst Pedersen and Marylesa Howard and Megan Owen and Jens Petersen and Jie Shi and Qiuping Xu},
  title = {{Quantification and visualization of variation in anatomical trees}},
  year = {2014},
  eprint = {1410.2466},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1410.2466},
}
Figure