Interactive Co-Registration for Multi-Modal Cancer Imaging Data based on Segmentation Masks

Information

• Publication Type: Master Thesis
• Workgroup(s)/Project(s):
  • Visualization Group
• Date: May 2022
• Date (Start): 3. February 2021
• Date (End): 17. May 2022
• Diploma Examination: 27. June 2022
• Open Access: yes
• First Supervisor: Eduard Gröller
• Pages: 126
• Keywords: Multi-Modal Segmentation, Registration

Abstract

Endometrial cancer is the most common and most lethal gynecologic malignancy world-wide. Multiple MRI sequences are acquired per patient in gynecologic cancer research because they reveal different tissue characteristics. Radiomic tumor profiling extracts features from medical imaging data aiming to find new tumor imaging biomarkers. Co-registration and tumor segmentation of multi-sequential MRI data build the base for radiomic tumor profiling. Many approaches exist that aim to automate these time-consuming manual processes. After automatic co-registration, volumes are often still misaligned. This lack of registration quality has an impact on the results of radiomic tumor profiling, since we cannot ensure voxel integrity. We distinguish between rigid and deformable registration. Rigid registration transforms a volume using only translation and rotation parameters, while deformable registration can include local deformations. Tumors are rigid structures compared to the tissue around them. Therefore, rigid co-registration can be sufficient to align tumors. However, to analyze also surrounding structures, deformable registration is necessary. Even though tumors are rigid structures, they can appear slightly different in the varying sequences due to imaging physics. Applying deformable registration to the whole image can result in tumor deformations that do not resemble the underlying biological tissue characteristics and can alter important information about tumor tissue characteristics. To address these two problems, we propose the web-based application MuSIC (Multi-Sequential Interactive Co-registration). The tool allows medical experts to co-register multiple sequences simultaneously based on a pre-defined segmentation mask that has been generated for one of the sequences. In our workflow, a simulated-annealing-based shape matching algorithm searches for the tumor position in each sequence that can vary in translation and rotation parameters. We present the updated segmentation positions to the user, who can interactively adapt the positions if needed. We include multi-modal visualization techniques for visual quality assessment during this procedure. Based on the positioning of the segmentation masks, we register the volumes. We allow for both rigid and deformable co-registration. Due to our approach based on segmentation masks, we apply local transformations mainly outside the tumor tissue in deformable registration. We evaluate our approach in a usability analysis with medical and machine learning experts. They find the tool very intuitive and especially the medical experts clearly see themselves using MuSIC in the future.

Additional Files and Images

Weblinks

• Entry in reposiTUm (TU Wien Publication Database)
• DOI: 10.34726/hss.2022.96562

BibTeX

@mastersthesis{Eichner2022,
  title =      "Interactive Co-Registration for Multi-Modal Cancer Imaging
               Data based on Segmentation Masks",
  author =     "Tanja Eichner",
  year =       "2022",
  abstract =   "Endometrial cancer is the most common and most lethal
               gynecologic malignancy world-wide. Multiple MRI sequences
               are acquired per patient in gynecologic cancer research
               because they reveal different tissue characteristics.
               Radiomic tumor profiling extracts features from medical
               imaging data aiming to find new tumor imaging biomarkers.
               Co-registration and tumor segmentation of multi-sequential
               MRI data build the base for radiomic tumor profiling. Many
               approaches exist that aim to automate these time-consuming
               manual processes. After automatic co-registration, volumes
               are often still misaligned. This lack of registration
               quality has an impact on the results of radiomic tumor
               profiling, since we cannot ensure voxel integrity. We
               distinguish between rigid and deformable registration. Rigid
               registration transforms a volume using only translation and
               rotation parameters, while deformable registration can
               include local deformations. Tumors are rigid structures
               compared to the tissue around them. Therefore, rigid
               co-registration can be sufficient to align tumors. However,
               to analyze also surrounding structures, deformable
               registration is necessary. Even though tumors are rigid
               structures, they can appear slightly different in the
               varying sequences due to imaging physics. Applying
               deformable registration to the whole image can result in
               tumor deformations that do not resemble the underlying
               biological tissue characteristics and can alter important
               information about tumor tissue characteristics. To address
               these two problems, we propose the web-based application
               MuSIC (Multi-Sequential Interactive Co-registration). The
               tool allows medical experts to co-register multiple
               sequences simultaneously based on a pre-defined
               segmentation mask that has been generated for one of the
               sequences. In our workflow, a simulated-annealing-based
               shape matching algorithm searches for the tumor position in
               each sequence that can vary in translation and rotation
               parameters. We present the updated segmentation positions to
               the user, who can interactively adapt the positions if
               needed. We include multi-modal visualization techniques for
               visual quality assessment during this procedure. Based on
               the positioning of the segmentation masks, we register the
               volumes. We allow for both rigid and deformable
               co-registration. Due to our approach based on segmentation
               masks, we apply local transformations mainly outside the
               tumor tissue in deformable registration. We evaluate our
               approach in a usability analysis with medical and machine
               learning experts. They find the tool very intuitive and
               especially the medical experts clearly see themselves using
               MuSIC in the future.",
  month =      may,
  pages =      "126",
  address =    "Favoritenstrasse 9-11/E193-02, A-1040 Vienna, Austria",
  school =     "Research Unit of Computer Graphics, Institute of Visual
               Computing and Human-Centered Technology, Faculty of
               Informatics, TU Wien",
  keywords =   "Multi-Modal Segmentation, Registration",
  URL =        "https://www.cg.tuwien.ac.at/research/publications/2022/Eichner2022/",
}