Automatic bone density evaluation from CT images

Serena Bonaretti, Philippe Büchler, Nils Reimers, W Schmidt, Christof Seiler, Stefan Weber, Mauricio ReyesOpen 2015 · DOI

TL;DR

This paper introduces an automated method to quantify bone density from CT scans, differentiating between cortical bone, trabecular bone, and marrow. The approach aims to provide a more detailed understanding of bone health, potentially aiding in osteoporosis prediction and research into biomechanical properties.

What Problem Does This Paper Address?

Bone density is a critical indicator for predicting osteoporosis and fracture risk, traditionally assessed using methods like DEXA. However, DEXA primarily focuses on specific bone regions and may not capture the nuanced density variations within different bone tissues. In biomedical research, precise bone density information is crucial for applications like creating subject-specific finite element models or evaluating bone-implant fit. This paper addresses the need for a more automated and detailed method to assess bone density directly from CT images, specifically by analyzing density differences across distinct histological regions (cortical bone, trabecular bone, marrow) and exploring variations between sexes and sides of the body.

Methods

The authors developed an automated pipeline for bone density evaluation from CT images. The core of their method involves:

  • Image Registration: A two-stage registration process was employed for each bone group (femur left/right, tibia left/right).
  • Rigid Registration: This step aligned all bones to a common coordinate system.
  • Non-Rigid Registration: Using the Diffeomorphic Demons algorithm, this stage established anatomical correspondences between voxels, allowing for precise mapping of regions across different scans.
  • Histological Region Segmentation: For each bone, reference images were manually segmented into three key histological regions: cortical bone, trabecular bone, and marrow. The segmentation from these reference images was then transferred to other bones in the dataset using the established correspondences.
  • Density Quantification: For each of the three histological regions within each bone, the mean and median of the grey level values (representing density) were extracted. The mean was used to capture the general density distribution, while the median was employed to mitigate the impact of outliers.
  • Statistical Analysis: Differences in bone density were statistically tested to compare groups (male vs. female) and sides (left vs. right).

The entire image processing pipeline was implemented using C++ with ITK and VTK libraries, and statistical analysis was performed in R.

Datasets

The dataset comprised CT images of left and right femurs and tibias from Caucasian males and females. The paper provides summary statistics for the mean age and number of patients for each group:

  • Femur:
    • Left Male: Age 65.9±14.3, 69 patients
    • Right Male: Age 65.8±14.4, 66 patients
    • Left Female: Age 64.7±16.4, 86 patients
    • Right Female: Age 63.8±16.5, 81 patients
  • Tibia:
    • Left Male: Age 65.9±15.4, 65 patients
    • Right Male: Age 66.4±14.7, 61 patients
    • Left Female: Age 64.1±16.6, 80 patients
    • Right Female: Age 64.7±16.4, 78 patients

The images were manually segmented prior to their use in the study.

Key Insights and Ideas

  • Automated Multi-Tissue Density Analysis: The paper moves beyond single-value bone density metrics by automatically segmenting and quantifying density in distinct histological regions (cortical, trabecular, marrow) from CT scans. This offers a more granular understanding of bone composition.
  • Sex-Based Density Differences: Preliminary results suggest that males generally exhibit higher bone density across all three histological regions compared to females. This aligns with general physiological understanding but is quantified here in a novel, automated manner.
  • Inter-Sex Variability: The study observed that female bone densities were more spread out than male densities, hinting at potentially greater individual variability or factors influencing bone density in women.
  • Cortical Bone Density Discrepancy: The measured cortical bone densities were lower than expected. The authors hypothesize this could be due to challenges in accurately segmenting the thin cortical layer, suggesting potential limitations in the registration or segmentation accuracy for this specific tissue.

Key Results

  • Cortical Bone Density: Mean and median values for cortical bone density were found to be lower than expected.
  • Trabecular Bone and Marrow Density: Densities for trabecular bone and marrow were within the expected ranges.
  • Sex Differences: Males showed significantly higher bone density than females across all three histological regions (p<0.05).
  • Density Variability: Female bone densities exhibited greater spread (variability) compared to males.
  • Side Differences: No significant differences in bone density were found between the left and right sides of the body for either males or females.

Limitations and Future Work

  • Cortical Bone Accuracy: The authors acknowledge that the lower-than-expected cortical bone density measurements might be linked to errors in correspondence detection, particularly due to the small thickness of the cortical bone. This suggests a need for improved methods to accurately segment and quantify density in very thin structures.
  • Age-Related Variations: While the dataset included a range of ages, the discussion briefly touches upon the variation of bone density affecting women after 40 years old. Future work could delve deeper into age-specific density changes and their correlation with the observed sex differences.
  • Clinical Validation: The paper presents preliminary results. Further validation with larger, more diverse datasets and comparison against established clinical measures would be beneficial to confirm the clinical utility of this automated technique for osteoporosis prediction or fracture risk assessment.

Summary updated

Abstract

Introduction The evaluation of the bone density from 2D and 3D medical images is an important task both in clinical and research fields. Clinically it is considered one of the parameters of prediction of osteoporosis and therefore bone fracture risk. It is commonly measured using the Dual Energy Xray Absorptiometry (DEXA), based on X-ray images of vertebrae or proximal femur and it allows the classification of the patient as normal, affected by osteopenia, or by osteoporosis. In biomedical research, the density of the bone represents an important parameter for example in those studies in which the exact knowledge of the mechanical properties of the bone are involved, such as the creation of subject-specific finite elements models or on evaluating bone-implant fitting assessment. The density changes inside the bone with relationship to the region of interest. From the histological point of view, the long bones such as femur ad tibia can be considered composed by three main parts. The outer layer is made by cortical bone, the middle layer is composed by trabecular bone and the inner layer is represented by the marrow. The bone density decreases passing form the outer to the inner part of the bone. In this paper we propose a method to evaluate automatically the bone density from CT images. The method was used in order to analyze possible differences between male and female bone density and the variations between the left and the right sides of the body.
BibTeX
@inproceedings{bonaretti2015automatic,
  author = {Serena Bonaretti and Philippe Büchler and Nils Reimers and W Schmidt and Christof Seiler and Stefan Weber and Mauricio Reyes},
  title = {{Automatic bone density evaluation from CT images}},
  booktitle = {Open},
  year = {2015},
  doi = {10.7892/boris.30520},
}
Figure