Enhancing the resolution of brain MRIs using deep learning.
(2024) In Master’s Theses in Mathematical Sciences FMAM05 20241Mathematics (Faculty of Engineering)
- Abstract
- Magnetic resonance imaging (MRI) is a medical imaging technique used to generate 3D cross-sections of body parts for medical diagnosis and monitoring. MRI scanners have a resolution that depends on the strength of their magnetic field, measured in Teslas (T). Most clinical scanners have a resolution of 1.5T and 3T resolution is used in wealthy hospitals and for research. A few research centers also have a 7T
scanner, a resolution only ever used for research. Yet, 7T images offer a very detailed view of the body, which helps researching and diagnosing diseases.
Deep learning models can be used to create super resolution models that generate a synthetic 7T MRI scan from a 3T MRI scan. The generated images carry information of a higher... (More) - Magnetic resonance imaging (MRI) is a medical imaging technique used to generate 3D cross-sections of body parts for medical diagnosis and monitoring. MRI scanners have a resolution that depends on the strength of their magnetic field, measured in Teslas (T). Most clinical scanners have a resolution of 1.5T and 3T resolution is used in wealthy hospitals and for research. A few research centers also have a 7T
scanner, a resolution only ever used for research. Yet, 7T images offer a very detailed view of the body, which helps researching and diagnosing diseases.
Deep learning models can be used to create super resolution models that generate a synthetic 7T MRI scan from a 3T MRI scan. The generated images carry information of a higher resolution and display very tiny but important brain parts, otherwise invisible on the 3T. The creation of such a model relies on training it on pairs composed of a 3T and a 7T brain scan acquired from the same patient.
In this work, we present how we handled the preprocessing of our 7T and 3T images pairs, as well as the augmentation methods that can be used to enhance the diversity of the dataset. Then, we introduce our deep learning model, an attention residual U-net with an extra layer specific to super resolution. We also explain how to build a generative adversarial network using this model as a generator. We then discuss how we used the models on the data, the parameters of the models and how we can assess the quality of the results. Finally, we display the results of three models, one trained on a simple but small dataset, two trained on a lower quality but bigger dataset, one being a U-net and the other a GAN U-net. We discuss our achievements and the limitations that negatively affected the results, as well as what could be done to improve them. (Less) - Popular Abstract
- Medical scanners are widely used for patient care and for research, as they take 3D pictures highlighting body properties. With the recent developments in AI, new medical image processing tools have been created and it has become crucial to keep developing them to improve patient's health.
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9174214
- author
- Gicquel, Malo LU
- supervisor
-
- Gabrielle Flood LU
- Jacob Vogel LU
- organization
- course
- FMAM05 20241
- year
- 2024
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- Deep learning, MRI, magnetic resonance imaging, machine learning, CNN, GAN, U-net, brain MRI, image analysis, MRI processing.
- publication/series
- Master’s Theses in Mathematical Sciences
- report number
- LUTFMA-3558-2024
- ISSN
- 1404-6342
- other publication id
- 2024:E67
- language
- English
- id
- 9174214
- date added to LUP
- 2024-09-23 14:34:16
- date last changed
- 2024-09-23 14:34:16
@misc{9174214,
abstract = {{Magnetic resonance imaging (MRI) is a medical imaging technique used to generate 3D cross-sections of body parts for medical diagnosis and monitoring. MRI scanners have a resolution that depends on the strength of their magnetic field, measured in Teslas (T). Most clinical scanners have a resolution of 1.5T and 3T resolution is used in wealthy hospitals and for research. A few research centers also have a 7T
scanner, a resolution only ever used for research. Yet, 7T images offer a very detailed view of the body, which helps researching and diagnosing diseases.
Deep learning models can be used to create super resolution models that generate a synthetic 7T MRI scan from a 3T MRI scan. The generated images carry information of a higher resolution and display very tiny but important brain parts, otherwise invisible on the 3T. The creation of such a model relies on training it on pairs composed of a 3T and a 7T brain scan acquired from the same patient.
In this work, we present how we handled the preprocessing of our 7T and 3T images pairs, as well as the augmentation methods that can be used to enhance the diversity of the dataset. Then, we introduce our deep learning model, an attention residual U-net with an extra layer specific to super resolution. We also explain how to build a generative adversarial network using this model as a generator. We then discuss how we used the models on the data, the parameters of the models and how we can assess the quality of the results. Finally, we display the results of three models, one trained on a simple but small dataset, two trained on a lower quality but bigger dataset, one being a U-net and the other a GAN U-net. We discuss our achievements and the limitations that negatively affected the results, as well as what could be done to improve them.}},
author = {{Gicquel, Malo}},
issn = {{1404-6342}},
language = {{eng}},
note = {{Student Paper}},
series = {{Master’s Theses in Mathematical Sciences}},
title = {{Enhancing the resolution of brain MRIs using deep learning.}},
year = {{2024}},
}