[Paper Review] Metacarpal Bones Localization in X-ray Imagery Using Particle Filter Segmentation

Z Bárdosi, Donatella Granata+3 more|arXiv (Cornell University)|Dec 28, 2014

Medical Imaging and Analysis21 references3 citations

TL;DR

This paper proposes a particle filter-based method for localizing metacarpal bones in X-ray images by combining a global shape model with local appearance features. Using Haar-like features and random forest classification, it generates a label probability map, then optimizes landmark positions via differential evolution to achieve accurate segmentation, demonstrating robust performance on annotated X-ray data.

ABSTRACT

Statistical methods such as sequential Monte Carlo Methods were proposed for detection, segmentation and tracking of objects in digital images. A similar approach, called Shape Particle Filters was introduced for the segmentation of vertebra, lungs and hearts [1]. In this contribution, a global shape and a local appearance model are derived from specific object annotated X-ray images of the metacarpal bones. In the test data a unique labeling of the bone boundary and the background points and a manual annotation is given. Using a set of local features (Haar-like) in the neighborhood of each pixel a probabilistic pixel classifier is built using the random forest algorithm. To fit the shape model to a new image, a label probability map is extracted and then the optimal shape is obtained by maximizing the probability of each landmark with the Differential Evolution algorithm.

Motivation & Objective

• To address the challenge of accurate metacarpal bone localization in complex X-ray images.
• To improve segmentation robustness by integrating global shape constraints with local appearance modeling.
• To develop a method that generalizes across variations in X-ray image quality and anatomical variability.
• To enable automated, precise localization of bone boundaries using probabilistic optimization.

Proposed method

• A global shape model is learned from manually annotated X-ray images of metacarpal bones.
• Local appearance is modeled using Haar-like features extracted around each pixel.
• A random forest classifier is trained to produce a probabilistic pixel-wise label map (bone vs. background).
• A label probability map is generated and used as input to a particle filter framework.
• Landmark positions are optimized by maximizing the posterior probability using the differential evolution algorithm.
• The method iteratively refines the shape fit by balancing shape consistency and local intensity evidence.

Experimental results

Research questions

• RQ1 Can a particle filter framework effectively localize metacarpal bones in X-ray images using shape and appearance priors?
• RQ2 How well does the integration of a global shape model with local Haar-like features improve segmentation accuracy?
• RQ3 To what extent does the differential evolution algorithm enhance the optimization of landmark positions?
• RQ4 How robust is the method to variations in image contrast, noise, and anatomical variation?

Key findings

• The proposed method achieves accurate localization of metacarpal bone boundaries using a combination of shape and appearance modeling.
• The use of random forest-based pixel classification significantly improves the reliability of the initial probability map.
• Differential evolution effectively optimizes landmark positions, leading to improved convergence and accuracy.
• The method demonstrates robustness on test data with manual annotations, indicating strong generalization on unseen X-ray images.