Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Domain Adaptation for Combined CT and CBCT Deep Learning Segmentation

Berg, Jonas LU (2021) In Master’s Theses in Mathematical Sciences FMAM05 20211
Mathematics (Faculty of Engineering)
Abstract
Computed tomography (CT) segmentation models are frequently used within radiotherapy treatment planning, but similar models are not available to the related imaging modality cone beam computed tomography (CBCT) due to the scarcity of labeled data from this domain. Such models could have multiple clinical applications whereby it is of interest to study whether the CT segmentation models can be adapted to generalize to the CBCT domain. This thesis applies multiple different domain adaptation techniques to the male pelvic segmentation problem and compares the relative performance of the models for both CT and CBCT segmentation. The results indicate that all domain adaptation techniques yield large improvements compared to the baseline results... (More)
Computed tomography (CT) segmentation models are frequently used within radiotherapy treatment planning, but similar models are not available to the related imaging modality cone beam computed tomography (CBCT) due to the scarcity of labeled data from this domain. Such models could have multiple clinical applications whereby it is of interest to study whether the CT segmentation models can be adapted to generalize to the CBCT domain. This thesis applies multiple different domain adaptation techniques to the male pelvic segmentation problem and compares the relative performance of the models for both CT and CBCT segmentation. The results indicate that all domain adaptation techniques yield large improvements compared to the baseline results and deformable image registration (DIR), but that a novel data augmentation pipeline suggested in this work might be the most efficient route to solving the domain shift problem. This data augmentation pipeline notably improves CBCT dice similarity coefficient (DSC) scores for bladder to 0.900 (baseline 0.553) and for rectum to 0.850 (baseline 0.605). The CBCT results obtained for this method are comparable to the baseline performance on the CT domain, indicating that the data augmentation approach comes close to completely solving the studied domain shift problem. (Less)
Please use this url to cite or link to this publication:
author
Berg, Jonas LU
supervisor
organization
course
FMAM05 20211
year
type
H2 - Master's Degree (Two Years)
subject
keywords
CT, CBCT, Deep Learning, Medical Image Segmentation, Radiotherapy, Domain Adaptation, CycleGAN, Domain Adversarial Neural Network, U-Net, Data Augmentation, Machine Learning
publication/series
Master’s Theses in Mathematical Sciences
report number
LUTFMA-3451-2021
ISSN
1404-6342
other publication id
2021:E34
language
English
id
9060317
date added to LUP
2021-07-14 14:35:46
date last changed
2021-07-14 14:35:46
@misc{9060317,
  abstract     = {{Computed tomography (CT) segmentation models are frequently used within radiotherapy treatment planning, but similar models are not available to the related imaging modality cone beam computed tomography (CBCT) due to the scarcity of labeled data from this domain. Such models could have multiple clinical applications whereby it is of interest to study whether the CT segmentation models can be adapted to generalize to the CBCT domain. This thesis applies multiple different domain adaptation techniques to the male pelvic segmentation problem and compares the relative performance of the models for both CT and CBCT segmentation. The results indicate that all domain adaptation techniques yield large improvements compared to the baseline results and deformable image registration (DIR), but that a novel data augmentation pipeline suggested in this work might be the most efficient route to solving the domain shift problem. This data augmentation pipeline notably improves CBCT dice similarity coefficient (DSC) scores for bladder to 0.900 (baseline 0.553) and for rectum to 0.850 (baseline 0.605). The CBCT results obtained for this method are comparable to the baseline performance on the CT domain, indicating that the data augmentation approach comes close to completely solving the studied domain shift problem.}},
  author       = {{Berg, Jonas}},
  issn         = {{1404-6342}},
  language     = {{eng}},
  note         = {{Student Paper}},
  series       = {{Master’s Theses in Mathematical Sciences}},
  title        = {{Domain Adaptation for Combined CT and CBCT Deep Learning Segmentation}},
  year         = {{2021}},
}