Advanced

Deep Learning to Predict Hip Fracture Risk from Clinical DXA-images

Yip, Karin LU and Husein, Meral (2018) BMEM01 20182
Department of Biomedical Engineering
Abstract
Osteoporosis is a bone disease that is defined as low bone mineral density (BMD) and results in an increased risk of bone fracture. It is a serious public health problem that causes excess mortality and major economical and social impact. Today there are more than 8.9 million fractures connected to low bone mineral density worldwide. Hip fractures are of particular concern since they are associated with excess mortality as high as 18-33%. The bone mineral density can be assessed with a number of different techniques. One of those is dual energy X-ray absorptiometry (DXA) which also is the most widely used clinical tool.

An arising field with many possible applications is the field within artificial intelligence (AI). Within medicine,... (More)
Osteoporosis is a bone disease that is defined as low bone mineral density (BMD) and results in an increased risk of bone fracture. It is a serious public health problem that causes excess mortality and major economical and social impact. Today there are more than 8.9 million fractures connected to low bone mineral density worldwide. Hip fractures are of particular concern since they are associated with excess mortality as high as 18-33%. The bone mineral density can be assessed with a number of different techniques. One of those is dual energy X-ray absorptiometry (DXA) which also is the most widely used clinical tool.

An arising field with many possible applications is the field within artificial intelligence (AI). Within medicine, smarter tools to ease diagnostics are developed with the help of AI. In this thesis project, a subfield of AI – deep learning and artificial neural networks – is explored to investigate whether an artificial neural network would be able to predict the hip fracture risk. The predictions are based on finding features of DXA-images that might indicate
an increased risk for fractures. Two approaches of implementing artificial neural networks are carried out: developing a custom network and using transfer learning on a pre-trained ResNet network.

Different network architectures were investigated and developed. Experimentation involved both adding different network layers in multiple combinations and tuning of hyperparameters of the networks. The results from the networks were compared to the area under curve (AUC) obtained from the receiver operating characteristic (ROC) based on the BMD from the Malmö cohort. The AUC was 0.7497. The best AUC-value for the custom network was 0.7821, which is better than the AUC based on the BMD. The best AUC for the tuned ResNet-model was 0.6277, which is worse. The results in this thesis indicate that further exploration of using artificial neural networks as part of a diagnostic tool should be done. (Less)
Popular Abstract (Swedish)
Kan risk för benbrott spås med artificiell intelligens?

Höftfrakturer, framförallt benbrott av lårbenshalsen, är ofta ett resultat av benskörhet. Benbrotten drabbar främst äldre personer och leder till sämre livskvalitet; de är även förknippade med högre dödlighet, så hög som 18-33% första året. Även om benskörhet ökar risken för benbrott, är inte alla drabbade bensköra. För att kunna förhindra ett benbrott behöver de som har störst risk att drabbas upptäckas i tid.
Please use this url to cite or link to this publication:
author
Yip, Karin LU and Husein, Meral
supervisor
organization
course
BMEM01 20182
year
type
H2 - Master's Degree (Two Years)
subject
keywords
hip fracture, osteoporosis, deep learning, neural network, DXA
language
English
additional info
2018-20
id
8964114
date added to LUP
2019-01-14 14:00:27
date last changed
2019-01-14 14:00:27
@misc{8964114,
  abstract     = {Osteoporosis is a bone disease that is defined as low bone mineral density (BMD) and results in an increased risk of bone fracture. It is a serious public health problem that causes excess mortality and major economical and social impact. Today there are more than 8.9 million fractures connected to low bone mineral density worldwide. Hip fractures are of particular concern since they are associated with excess mortality as high as 18-33%. The bone mineral density can be assessed with a number of different techniques. One of those is dual energy X-ray absorptiometry (DXA) which also is the most widely used clinical tool.

An arising field with many possible applications is the field within artificial intelligence (AI). Within medicine, smarter tools to ease diagnostics are developed with the help of AI. In this thesis project, a subfield of AI – deep learning and artificial neural networks – is explored to investigate whether an artificial neural network would be able to predict the hip fracture risk. The predictions are based on finding features of DXA-images that might indicate
an increased risk for fractures. Two approaches of implementing artificial neural networks are carried out: developing a custom network and using transfer learning on a pre-trained ResNet network.

Different network architectures were investigated and developed. Experimentation involved both adding different network layers in multiple combinations and tuning of hyperparameters of the networks. The results from the networks were compared to the area under curve (AUC) obtained from the receiver operating characteristic (ROC) based on the BMD from the Malmö cohort. The AUC was 0.7497. The best AUC-value for the custom network was 0.7821, which is better than the AUC based on the BMD. The best AUC for the tuned ResNet-model was 0.6277, which is worse. The results in this thesis indicate that further exploration of using artificial neural networks as part of a diagnostic tool should be done.},
  author       = {Yip, Karin and Husein, Meral},
  keyword      = {hip fracture,osteoporosis,deep learning,neural network,DXA},
  language     = {eng},
  note         = {Student Paper},
  title        = {Deep Learning to Predict Hip Fracture Risk from Clinical DXA-images},
  year         = {2018},
}