Hur U-Net hyperparametrar och datakvalitet påverkar minskning av bildstörningar i magnetresonanstomografi
(2025) EEML05 20251Department of Biomedical Engineering
- Abstract
- Medical imaging plays a vital role in modern diagnostics, enabling accurate disease detection and treatment planning. In recent years, deep learning-especially using convolutional neural networks (CNNs) has revolutionized medical image segmentation. One prominent CNN architecture is U-net, developed specifically for biomedical image segmentation. Its structure with symmetric encoder-decoder layers and skip connections enables the model to perform well even when only a small amount of training data is available. In this study, we investigate how the choice of U-net settings (hyperparameters) and data quality impact the model's ability to reconstruct diffusion-weighted MR images (DWI) from anatomical T1-weighted scans. MRI images often... (More)
- Medical imaging plays a vital role in modern diagnostics, enabling accurate disease detection and treatment planning. In recent years, deep learning-especially using convolutional neural networks (CNNs) has revolutionized medical image segmentation. One prominent CNN architecture is U-net, developed specifically for biomedical image segmentation. Its structure with symmetric encoder-decoder layers and skip connections enables the model to perform well even when only a small amount of training data is available. In this study, we investigate how the choice of U-net settings (hyperparameters) and data quality impact the model's ability to reconstruct diffusion-weighted MR images (DWI) from anatomical T1-weighted scans. MRI images often contain artifacts—errors caused by movement, metal objects, or machine settings. These artifacts can decrease model performance if present in training data. We manually curated the dataset to separate artifact-free images from noisy ones, using a custom MATLAB GUI for binary classification. The U-net was trained multiple times with varying hyperparameters such as batch size, learning rate, patch size, and network depth, following methodological guidance from recent literature. The performance of each model was evaluated visually, with particular focus on the reconstruction of critical brain structures. Our results show that models trained on unfiltered data are prone to reproducing artifacts, particularly with prolonged training. In contrast, training on curated data led to more consistent anatomical reconstructions and reduced the model’s sensitivity to hyperparameter variation. We also found that simplified networks with reduced depth or features underperformed in structural accuracy. This study highlights the importance of data quality and parameter optimization in medical image synthesis using deep learning. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9201852
- author
- Fridberg, Filip LU and Sibley, Matilda LU
- supervisor
- organization
- alternative title
- Impact of U-Net Hyperparameters and Data Quality on Artifact Reduction in Magnetic Resonance Imaging
- course
- EEML05 20251
- year
- 2025
- type
- M2 - Bachelor Degree
- subject
- language
- Swedish
- id
- 9201852
- date added to LUP
- 2025-07-01 09:23:14
- date last changed
- 2025-07-01 09:23:14
@misc{9201852,
abstract = {{Medical imaging plays a vital role in modern diagnostics, enabling accurate disease detection and treatment planning. In recent years, deep learning-especially using convolutional neural networks (CNNs) has revolutionized medical image segmentation. One prominent CNN architecture is U-net, developed specifically for biomedical image segmentation. Its structure with symmetric encoder-decoder layers and skip connections enables the model to perform well even when only a small amount of training data is available. In this study, we investigate how the choice of U-net settings (hyperparameters) and data quality impact the model's ability to reconstruct diffusion-weighted MR images (DWI) from anatomical T1-weighted scans. MRI images often contain artifacts—errors caused by movement, metal objects, or machine settings. These artifacts can decrease model performance if present in training data. We manually curated the dataset to separate artifact-free images from noisy ones, using a custom MATLAB GUI for binary classification. The U-net was trained multiple times with varying hyperparameters such as batch size, learning rate, patch size, and network depth, following methodological guidance from recent literature. The performance of each model was evaluated visually, with particular focus on the reconstruction of critical brain structures. Our results show that models trained on unfiltered data are prone to reproducing artifacts, particularly with prolonged training. In contrast, training on curated data led to more consistent anatomical reconstructions and reduced the model’s sensitivity to hyperparameter variation. We also found that simplified networks with reduced depth or features underperformed in structural accuracy. This study highlights the importance of data quality and parameter optimization in medical image synthesis using deep learning.}},
author = {{Fridberg, Filip and Sibley, Matilda}},
language = {{swe}},
note = {{Student Paper}},
title = {{Hur U-Net hyperparametrar och datakvalitet påverkar minskning av bildstörningar i magnetresonanstomografi}},
year = {{2025}},
}