uu.seUppsala University Publications
Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Automatic Recognition of Abdominal Organs in Whole-Body Water Fat MRI
2017 (English)Independent thesis Advanced level (professional degree), 20 credits / 30 HE creditsStudent thesis
Abstract [en]

As imaging has become part of the clinical routine in medicine, automatic analysis of medical images has gained increasing interest and importance. Imiomics, developed at the department of radiology, Uppsala universitet, is an application for automatic analysis of whole-body magnetic resonance (MR) images and positron emission tomography (PET) images. Imiomics gives insight into questions such as correlation between genetics and physical morphology and diseases, and could be used for following patients before and after medical treatment. Imiomics is based on image registration, which relates to matching of different image data, but registration often fails in body regions with high variability, such as the abdominal organs, which have large variations in size and shape between different subjects.

One way of improving the registration in difficult regions is by localization of anatomical structures beforehand. However, manual localization and segmentation are often time-consuming and also subjective procedures. To this end, this thesis investigated whether machine learning could be used for automatic recognition and localization of abdominal organs from whole-body MR images by building up a computer vision system. This work aimed for recognition of a few organs in the abdomen and torso; liver, kidney, heart, spine, stomach, and fat tissue. The data set consisted of 10 subjects, where seven of them where used for training and three subjects were used for validation of the model. The algorithm chosen for the task was random forest and the computational software used was MATLAB.

Expressive texture features were determined during a training phase by filtering the images with various kernels and by calculation of co-occurrence matrices. Also, features based on spatial position and distance were calculated. A large number of feature was employed as the baseline approach. However, the dimension of the feature space was reduced to limit computational needs. Dimension reduction was applied in order to select the most important features for the recognition task. Some experiments of feature selection were tried such as filter methods, and sequential forward feature selection, also some experiments with random forest feature importance were done. Sequential forward feature selection reduced the number of feature the most without losing predictive power of any considerable amount. The selected features were often related to position and distances, which also have low computational cost. In order to invoke more of those feature, a second layer of random forest was introduced. The first layer of classifier produced probability maps, from which estimated center of mass of the regions of interest were extracted. As a result, coordinates and distances relative these landmarks of estimated center of mass were used as features for the second layer of random forest classifier.

In total, four variations of the classifier design were tested and compared; the baseline feature set, a reduced feature set by using feature selection, a two-layer classifier design, and a two-layer design with a second feature selection applied. The most successful design was the two layer design of random forest, which achieved an average error on the estimated of center of mass with 1.4 cm, without any post-processing applied, within an average run time of two minutes for classification of a test volume of the torso. However, there was no large difference in predictive power between the different classifier designs.

The computer vision pipeline constructed reached reasonable performance in the localization task, and machine learning algorithms such as random forest could successfully be used to localize anatomical structures, even with such a small data set as 10 subjects. Thus, such a system is considered to be useful as a pre-step for Imiomics. 

Place, publisher, year, edition, pages
2017.
Series
UPTEC F, ISSN 1401-5757 ; 17036
National Category
Engineering and Technology
Identifiers
URN: urn:nbn:se:uu:diva-326814OAI: oai:DiVA.org:uu-326814DiVA, id: diva2:1128862
Educational program
Master Programme in Engineering Physics
Supervisors
Examiners
Available from: 2017-08-08 Created: 2017-07-31 Last updated: 2017-08-08Bibliographically approved

Open Access in DiVA

No full text in DiVA

Engineering and Technology

Search outside of DiVA

GoogleGoogle Scholar

urn-nbn

Altmetric score

urn-nbn
Total: 332 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf