In the era of global aging, hip fragility fracture has become a health problem. Its main cause is the phenomena of osteoporosis (OP). OP is a systemic illness of bone tissue characterized by low mineral density and deterioration of microarchitectural organization, inducing a higher risk of fractures, mainly in hip and spine. With this in mind, early prediction of hip fragility fractures would play a leading role, both in reducing public health costs and saving human lives. The purpose of this project is to discriminate between patients with high-level or low-level risk of hip fragility fracture. To accomplish this study, deep learning (DL) was exploited, investigating the performance and behavior of different convolutional neural network (CNN) architectures in this binary classification task. The provided dataset, constituted of lumbar computed tomography (CT) scans of 142 patients, is coming from two different radiology departments in Stockholm: Karolinska University Hospital in Flemingsberg and in Solna. For each patient, two types of image were considered: lumbar axial and lumbar sagittal viewpoints. For the scope of this thesis, the two selected CNNs were VGG16 and ResNet50. The networks were pretrained on ImageNet dataset and transfer learning techniques were adopted to extract informative features from the training set. In order to obtain a model with higher generalization capability, ensemble modelling between VGG16 and ResNet50 was adopted. The classification accuracy was evaluated using 5-fold cross-validation. With the performed experiments, we achieved an area under the receiver operating characteristic curve (AUC) of 61.5% and 62.7% with ensemble modelling for axial and sagittal images, respectively. With the aim to adopt a more holistic approach, predictions from axial and sagittal perspectives were merged to obtain a single subject prediction, with a final AUC of 62.5%. To further validate our study case, class activation mapping (CAM) was exploited in order to point out the discriminative regions leading the networks to classify. Many of the results were only slightly better than 50%, this means the performed experiments show only slightly better performance than random guessing. For this reason, even if results are around 60% in terms of AUC, the outcomes given in this study are not conclusive and would require more research.

In the era of global aging, hip fragility fracture has become a health problem. Its main cause is the phenomena of osteoporosis (OP). OP is a systemic illness of bone tissue characterized by low mineral density and deterioration of microarchitectural organization, inducing a higher risk of fractures, mainly in hip and spine. With this in mind, early prediction of hip fragility fractures would play a leading role, both in reducing public health costs and saving human lives. The purpose of this project is to discriminate between patients with high-level or low-level risk of hip fragility fracture. To accomplish this study, deep learning (DL) was exploited, investigating the performance and behavior of different convolutional neural network (CNN) architectures in this binary classification task. The provided dataset, constituted of lumbar computed tomography (CT) scans of 142 patients, is coming from two different radiology departments in Stockholm: Karolinska University Hospital in Flemingsberg and in Solna. For each patient, two types of image were considered: lumbar axial and lumbar sagittal viewpoints. For the scope of this thesis, the two selected CNNs were VGG16 and ResNet50. The networks were pretrained on ImageNet dataset and transfer learning techniques were adopted to extract informative features from the training set. In order to obtain a model with higher generalization capability, ensemble modelling between VGG16 and ResNet50 was adopted. The classification accuracy was evaluated using 5-fold cross-validation. With the performed experiments, we achieved an area under the receiver operating characteristic curve (AUC) of 61.5% and 62.7% with ensemble modelling for axial and sagittal images, respectively. With the aim to adopt a more holistic approach, predictions from axial and sagittal perspectives were merged to obtain a single subject prediction, with a final AUC of 62.5%. To further validate our study case, class activation mapping (CAM) was exploited in order to point out the discriminative regions leading the networks to classify. Many of the results were only slightly better than 50%, this means the performed experiments show only slightly better performance than random guessing. For this reason, even if results are around 60% in terms of AUC, the outcomes given in this study are not conclusive and would require more research.

EARLY PREDICTION OF OSTEOPOROTIC HIP FRACTURES BY DEEP LEARNING ALGORITHMS

POZZA, GIACOMO
2021/2022

Abstract

In the era of global aging, hip fragility fracture has become a health problem. Its main cause is the phenomena of osteoporosis (OP). OP is a systemic illness of bone tissue characterized by low mineral density and deterioration of microarchitectural organization, inducing a higher risk of fractures, mainly in hip and spine. With this in mind, early prediction of hip fragility fractures would play a leading role, both in reducing public health costs and saving human lives. The purpose of this project is to discriminate between patients with high-level or low-level risk of hip fragility fracture. To accomplish this study, deep learning (DL) was exploited, investigating the performance and behavior of different convolutional neural network (CNN) architectures in this binary classification task. The provided dataset, constituted of lumbar computed tomography (CT) scans of 142 patients, is coming from two different radiology departments in Stockholm: Karolinska University Hospital in Flemingsberg and in Solna. For each patient, two types of image were considered: lumbar axial and lumbar sagittal viewpoints. For the scope of this thesis, the two selected CNNs were VGG16 and ResNet50. The networks were pretrained on ImageNet dataset and transfer learning techniques were adopted to extract informative features from the training set. In order to obtain a model with higher generalization capability, ensemble modelling between VGG16 and ResNet50 was adopted. The classification accuracy was evaluated using 5-fold cross-validation. With the performed experiments, we achieved an area under the receiver operating characteristic curve (AUC) of 61.5% and 62.7% with ensemble modelling for axial and sagittal images, respectively. With the aim to adopt a more holistic approach, predictions from axial and sagittal perspectives were merged to obtain a single subject prediction, with a final AUC of 62.5%. To further validate our study case, class activation mapping (CAM) was exploited in order to point out the discriminative regions leading the networks to classify. Many of the results were only slightly better than 50%, this means the performed experiments show only slightly better performance than random guessing. For this reason, even if results are around 60% in terms of AUC, the outcomes given in this study are not conclusive and would require more research.
2021
EARLY PREDICTION OF OSTEOPOROTIC HIP FRACTURES BY DEEP LEARNING ALGORITHMS
In the era of global aging, hip fragility fracture has become a health problem. Its main cause is the phenomena of osteoporosis (OP). OP is a systemic illness of bone tissue characterized by low mineral density and deterioration of microarchitectural organization, inducing a higher risk of fractures, mainly in hip and spine. With this in mind, early prediction of hip fragility fractures would play a leading role, both in reducing public health costs and saving human lives. The purpose of this project is to discriminate between patients with high-level or low-level risk of hip fragility fracture. To accomplish this study, deep learning (DL) was exploited, investigating the performance and behavior of different convolutional neural network (CNN) architectures in this binary classification task. The provided dataset, constituted of lumbar computed tomography (CT) scans of 142 patients, is coming from two different radiology departments in Stockholm: Karolinska University Hospital in Flemingsberg and in Solna. For each patient, two types of image were considered: lumbar axial and lumbar sagittal viewpoints. For the scope of this thesis, the two selected CNNs were VGG16 and ResNet50. The networks were pretrained on ImageNet dataset and transfer learning techniques were adopted to extract informative features from the training set. In order to obtain a model with higher generalization capability, ensemble modelling between VGG16 and ResNet50 was adopted. The classification accuracy was evaluated using 5-fold cross-validation. With the performed experiments, we achieved an area under the receiver operating characteristic curve (AUC) of 61.5% and 62.7% with ensemble modelling for axial and sagittal images, respectively. With the aim to adopt a more holistic approach, predictions from axial and sagittal perspectives were merged to obtain a single subject prediction, with a final AUC of 62.5%. To further validate our study case, class activation mapping (CAM) was exploited in order to point out the discriminative regions leading the networks to classify. Many of the results were only slightly better than 50%, this means the performed experiments show only slightly better performance than random guessing. For this reason, even if results are around 60% in terms of AUC, the outcomes given in this study are not conclusive and would require more research.
DEEP LEARNING
OSTEOPOROSIS
CLASSIFICATION
File in questo prodotto:
File Dimensione Formato  
Pozza_Giacomo.pdf

accesso riservato

Dimensione 3.94 MB
Formato Adobe PDF
3.94 MB Adobe PDF

The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.12608/30866