📊 statistics

Multivariate Planar Curves: A Statistical Framework for Shape Analysis in Images

This paper introduces a statistical framework for analyzing multivariate planar curves to jointly model multiple object contours in images, thereby preserving their relative geometric relationships and improving supervised classification accuracy, as demonstrated in cardiomegaly detection on chest X-rays.

Original authors: Issam-Ali Moindjié, Cédric Beaulac, Marie-Hélène Descary

Published 2026-07-24

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Issam-Ali Moindjié, Cédric Beaulac, Marie-Hélène Descary

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but instead of looking at fingerprints or footprints, you are looking at the outlines of things. In the world of computer vision and statistics, there is a field called shape analysis. Think of it as the art of describing an object not by its color or texture, but by its "skeleton" or "contour"—the line that traces its edge. Just like a silhouette can tell you if a person is tall and thin or short and round, these outlines carry a lot of information.

Usually, statisticians have been looking at just one outline at a time, like studying a single apple. But in the real world, objects rarely exist in isolation. A heart sits between two lungs; a car has wheels attached to a body. The paper you are about to read tackles a tricky question: How do we study the shapes of several objects together, as a single team, rather than as strangers standing next to each other? The authors argue that if you look at them separately, you miss the most important clues: how big one object is compared to another, how they are positioned relative to each other, and how they are oriented. This matters because in fields like medicine, the relationship between organs can be the key to diagnosing a disease.

The Story of the Shape-Shifting Team

Imagine you have a group of three friends: a Heart, a Left Lung, and a Right Lung. In a perfect world, they would always stand in the exact same spot, facing the same direction, holding hands in the same way. But in the real world, things get messy. Sometimes the photo is taken from a slightly different angle (rotation), sometimes the camera zooms in or out (scaling), and sometimes the friends shift their positions slightly (translation). Even worse, if you were to trace their outlines with a pen, you might start drawing the Heart at the top, but start drawing the Lungs at the bottom. This is called a "reparametrization" issue—it's like starting a song at the chorus instead of the beginning.

The authors of this paper, Issam-Ali Moindjié, Cédric Beaulac, and Marie-Hélène Descary, realized that previous methods treated these friends as if they were strangers. They would analyze the Heart's shape, then the Left Lung's shape, and then the Right Lung's shape, completely ignoring how they stood next to each other. The authors say, "Wait a minute! The relationship between them is the secret sauce."

To fix this, they invented a new statistical framework called Multivariate Planar Curves. Think of this as a magical backpack that ties the three friends together into a single, inseparable unit. Instead of looking at three separate outlines, the computer now sees one giant, complex shape made of three parts. This allows the computer to understand that if the Heart gets bigger, it might be squeezing the lungs, and that specific "squeeze" is a pattern worth noticing.

The Magic of "Alignment"

Before the computer can compare these shape-teams, it has to get them to stand in the same pose. This is the "alignment" problem. Imagine you have a photo of a dance crew. If one person is doing a handstand and another is standing on their head, you can't tell if they are doing the same dance move. You need to rotate and shift them so they are all standing upright and facing the same way.

The paper introduces a clever algorithm called Iterative Closest Function (ICF). It's like a game of "hot and cold" played by a computer. The computer guesses how much to rotate and shift the shapes to match a "template" (a perfect average version of the group). It checks the score, adjusts the guess, and repeats this process thousands of times until the shapes snap perfectly into place. The authors tested this on a fake dataset where they deliberately messed up the shapes with random rotations and shifts. The result? Their method was incredibly accurate, recovering the original shapes even when the noise was quite high. It's as if they could look at a blurry, spinning photo and perfectly reconstruct the original dance routine.

The Big Test: Spotting a Big Heart

The real test came when they applied this to a medical problem: cardiomegaly, which is a fancy word for an abnormally large heart. They used a dataset of chest X-rays where the lungs and heart had already been traced out (segmented). The goal was to teach a computer to tell the difference between a healthy person and someone with a big heart.

They set up a race between three different strategies:

The Naive Approach (CLA): Just feed the raw, messy outlines into the computer without fixing the rotation or starting points.
The Solo Approach (UNI): Fix the rotation for each organ separately, but treat them as three independent strangers.
The Team Approach (OUR): Use their new "Multivariate Planar Curve" method to fix the rotation for the whole group at once and treat them as a single unit.

The results were dramatic. When the data was already perfectly aligned, all three methods did okay. But when the authors deliberately messed up the data (simulating real-world messy X-rays), the Naive Approach collapsed, performing no better than random guessing (around 53-57% accuracy). It was completely confused by the rotation and shifting.

The Solo Approach did better, but it still missed the big picture. However, the Team Approach (OUR) remained rock-solid. It maintained high accuracy (around 80-86%) even when the data was messy. Why? Because by looking at the heart and lungs together, the computer could see the relative size and position. It didn't just see a "big heart"; it saw a "big heart relative to the lungs," which is the actual definition of the condition.

What This Means

The paper proves that when you have multiple objects in an image, you shouldn't just analyze them one by one. By tying them together into a single statistical object, you preserve the "inter-component information"—the clues hidden in how they relate to each other.

The authors are careful to note that this method relies on having the outlines already drawn (segmented). If the computer can't find the edges of the organs in the first place, this method can't work. They also suggest that while shapes are powerful, they might miss out on other clues like color. But for now, they have shown that treating a group of shapes as a single, coordinated team is a much smarter way to analyze images than treating them as a crowd of individuals. It's a reminder that sometimes, the whole is indeed greater than the sum of its parts.

Technical Summary: Multivariate Planar Curves: A Statistical Framework for Shape Analysis in Images

Problem Statement
The paper addresses the challenge of supervised classification in image analysis where the input data consists of segmented images containing multiple objects. While traditional statistical shape analysis focuses on single contours (univariate), many practical applications—such as medical diagnosis involving multiple anatomical structures—require the joint analysis of several objects. Existing approaches often treat these objects independently or rely on raw pixel data, potentially discarding critical geometric information regarding the relative position, scale, and orientation between objects. The authors aim to develop a statistical learning framework that models the joint shape of multiple objects as a single entity, preserving inter-component information while accounting for standard geometric deformations (translation, rotation, scaling, and reparametrization).

Methodology
The proposed framework extends Functional Data Analysis (FDA) and statistical shape analysis to multivariate planar curves.

Model Definition:
- An image containing $p$ objects is represented as a multivariate planar curve $\mathbf{C} = (\mathbf{C}_1^\top, \dots, \mathbf{C}_p^\top)^\top$ , where each $\mathbf{C}_j$ is a closed planar curve defined by coordinate functions $(X_j, Y_j)$ .
- The model assumes the observed curve is a deformed version of a latent shape $\tilde{\mathbf{C}}$ . The deformation includes global parameters shared across all contours (translation $\mathbf{T}$ , rotation $\theta$ , and scaling $\rho$ ) and local parameters specific to each contour (reparametrization $\gamma_{\delta_j}$ ).
- Crucially, unlike a naive extension where each contour has its own independent deformation parameters, this model treats relative variations between contours (e.g., the size of the heart relative to the lungs) as intrinsic characteristics of the multivariate shape, not as noise to be removed.
Estimation and Alignment:
- Functional Representation: Curves are represented in a finite Fourier basis to facilitate computation.
- Pre-shape Recovery: Translation and scaling are removed in closed form to obtain a "pre-shape."
- Joint Alignment: To recover the shape, the authors extend the Iterative Closest Function (ICF) algorithm. This alternating optimization procedure estimates the shared rotation angle $\theta$ $θ$ and individual reparametrization shifts $\delta_j$ $δ_{j}$ by minimizing the distance between the pre-shape and a template. The algorithm iterates between:
  - Estimating $\theta$ for fixed $\delta$ (solving a Procrustes problem across all $p$ components).
  - Estimating $\delta_j$ for fixed $\theta$ (solving $p$ independent univariate alignment problems).
Statistical Analysis:
- Shape Space: The space of shapes is defined as a quotient space under rotation and reparametrization.
- Fréchet Mean: A representative mean shape is estimated iteratively by aligning all pre-shapes to a current template and updating the template to the empirical mean.
- Tangent Space Linearization: To enable standard statistical methods, shapes are projected onto the tangent space at the Fréchet mean using the logarithm map.
- Classification: These tangent coordinates serve as predictors in functional classification models (e.g., functional linear models with logistic regression or linear discriminant analysis).

Key Contributions

Formal Definition: Introduction of the "multivariate planar curve" and a formal definition of multivariate shape that encodes relative geometric configurations between objects.
Unified Framework: Extension of core statistical shape analysis tools (alignment, dissimilarity measurement, Fréchet mean estimation, and tangent space representation) from single contours to multivariate curves.
Joint Alignment Algorithm: Development of an extended ICF algorithm capable of jointly estimating global rotation and individual reparametrizations for multiple contours.
Empirical Validation: Demonstration that jointly modeling contours preserves discriminative inter-component information that univariate approaches discard.

Results
The methodology was evaluated through two primary experiments:

Synthetic Data (Alignment Accuracy):
- A simulation study using 500 curves with varying noise levels ( $\sigma$ ) tested the recovery of rotation and reparametrization parameters.
- The cyclic mean squared error (cMSE) for both rotation and reparametrization remained low even under substantial noise, confirming the robustness of the joint alignment procedure.
Real-World Application (Cardiomegaly Detection):
- The framework was applied to a dataset of segmented chest X-rays (heart and two lungs) to detect cardiomegaly (an abnormally large heart).
- Scenario 1 (Pre-aligned data): The proposed multivariate approach (OUR) performed comparably to a classical naive approach (CLA) using raw curves, and better than a univariate approach (UNI) that treated organs independently.
- Scenario 2 (Misaligned data): When artificial misalignments were introduced, the naive approach (CLA) failed (accuracy dropped to ~53-57%, near random guessing). The shape-based methods remained robust. The proposed multivariate approach (OUR) significantly outperformed the univariate approach (UNI) by 3 to 11 percentage points across various classifiers (GL1, GL2, PLS, PCR).

Significance and Claims
The authors claim that their framework offers a parsimonious and interpretable alternative to pixel-based deep learning methods, particularly when segmented images are available. The primary significance lies in the ability to preserve inter-component information (relative position, scale, orientation) which is often essential for diagnosis but is lost when objects are analyzed separately.

The paper modestly notes that the approach relies on the quality of existing image segmentation. While the current method uses pre-segmented masks, the authors suggest future work could focus on extracting these curves directly from raw images and potentially integrating color information, though the current contribution is strictly limited to shape analysis of segmented objects. The results suggest that for tasks where the geometric relationship between multiple objects is the key discriminative factor, joint multivariate modeling is superior to both independent univariate analysis and naive raw-data approaches, especially in the presence of misalignment.

The Story of the Shape-Shifting Team

The Magic of "Alignment"

The Big Test: Spotting a Big Heart

What This Means

Technical Summary: Multivariate Planar Curves: A Statistical Framework for Shape Analysis in Images

More like this