Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Machine Learning Acceleration of the DXA2FEM Pipeline for Femoral Bone Strength Prediction

Ekebro, Joel LU and Sahlberg, Morgan LU (2026) BMEM01 20261
Division for Biomedical Engineering
Abstract
Fragility fractures are a major health concern and accurate estimation of bone strength is important for fracture risk assessment. Finite element (FE) models can provide reliable estimates of femoral bone strength, but traditional approaches typically require three-dimensional imaging and computationally intensive model generation. Previous research has enabled reconstruction of subject-specific FE models by only using dual-energy X-ray absorptiometry (DXA) images, but the optimization procedure used to estimate the model parameters can require several hours per patient, which is not feasible in a clinical scenario.

The aim of this thesis was to investigate whether machine learning can be used to accelerate these reconstruction... (More)
Fragility fractures are a major health concern and accurate estimation of bone strength is important for fracture risk assessment. Finite element (FE) models can provide reliable estimates of femoral bone strength, but traditional approaches typically require three-dimensional imaging and computationally intensive model generation. Previous research has enabled reconstruction of subject-specific FE models by only using dual-energy X-ray absorptiometry (DXA) images, but the optimization procedure used to estimate the model parameters can require several hours per patient, which is not feasible in a clinical scenario.

The aim of this thesis was to investigate whether machine learning can be used to accelerate these reconstruction pipelines. Two neural network architectures were developed: a baseline convolutional neural network (CNN) and a transfer learning network based on EfficientNetV2B2. The networks were trained using DXA images from the MrOS Sweden cohort to either predict the statistical shape and appearance model (SSAM) parameters used for reconstruction or predict bone strength directly from the images. Artificially generated digitally reconstructed radiographs were also evaluated as synthetic training data.

The results show that neural networks can predict SSAM parameters and approximate bone strength estimates with a substantially lower computational cost; potentially saving several hours in computation time. The baseline CNN achieved slightly lower prediction errors than the transfer learning model in most metrics, both predicting reconstruction parameters and bone strength. The most accurate bone strength results were obtained when predicting bone strength directly from DXA images. Although the predicted SSAM parameters likely cannot yet replace the optimization procedure used in existing methods, they could be used to initialize the optimization algorithm. Future work will investigate whether such initialization can reduce convergence time and thereby accelerate the reconstruction process, as well as whether direct bone strength predictions are sufficiently accurate for clinical use. (Less)
Popular Abstract
Can AI Speed Up Bone Strength Estimation from X-Rays?

Fragility fractures are common and often happen after minor falls or even normal daily activity. Around one in three women and one in six men will experience such a fracture during their lifetime. These injuries are usually caused by reduced bone strength that develops over time and often goes unnoticed until a fracture occurs. Improving how bone strength is assessed could therefore help identify people at risk earlier.
Please use this url to cite or link to this publication:
author
Ekebro, Joel LU and Sahlberg, Morgan LU
supervisor
organization
course
BMEM01 20261
year
type
H2 - Master's Degree (Two Years)
subject
language
English
additional info
2026-05
id
9225841
date added to LUP
2026-05-04 10:32:03
date last changed
2026-05-04 10:32:03
@misc{9225841,
  abstract     = {{Fragility fractures are a major health concern and accurate estimation of bone strength is important for fracture risk assessment. Finite element (FE) models can provide reliable estimates of femoral bone strength, but traditional approaches typically require three-dimensional imaging and computationally intensive model generation. Previous research has enabled reconstruction of subject-specific FE models by only using dual-energy X-ray absorptiometry (DXA) images, but the optimization procedure used to estimate the model parameters can require several hours per patient, which is not feasible in a clinical scenario.

The aim of this thesis was to investigate whether machine learning can be used to accelerate these reconstruction pipelines. Two neural network architectures were developed: a baseline convolutional neural network (CNN) and a transfer learning network based on EfficientNetV2B2. The networks were trained using DXA images from the MrOS Sweden cohort to either predict the statistical shape and appearance model (SSAM) parameters used for reconstruction or predict bone strength directly from the images. Artificially generated digitally reconstructed radiographs were also evaluated as synthetic training data.

The results show that neural networks can predict SSAM parameters and approximate bone strength estimates with a substantially lower computational cost; potentially saving several hours in computation time. The baseline CNN achieved slightly lower prediction errors than the transfer learning model in most metrics, both predicting reconstruction parameters and bone strength. The most accurate bone strength results were obtained when predicting bone strength directly from DXA images. Although the predicted SSAM parameters likely cannot yet replace the optimization procedure used in existing methods, they could be used to initialize the optimization algorithm. Future work will investigate whether such initialization can reduce convergence time and thereby accelerate the reconstruction process, as well as whether direct bone strength predictions are sufficiently accurate for clinical use.}},
  author       = {{Ekebro, Joel and Sahlberg, Morgan}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Machine Learning Acceleration of the DXA2FEM Pipeline for Femoral Bone Strength Prediction}},
  year         = {{2026}},
}