Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
Machine learning for semantic segmentation of biomedical imaging data: ensemble 2D UNets for CT and PET images.
Thesis (MEng)--Stellenbosch University, 2021.
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | en_ZA |
| Published: |
Stellenbosch : Stellenbosch University
2021
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867613820475670528 |
|---|---|
| access_status_str | Open Access |
| author | Engelbrecht, J. A. |
| author2 | Nieuwoudt, M. J. |
| author_browse | Engelbrecht, J. A. Nieuwoudt, M. J. |
| author_facet | Nieuwoudt, M. J. Engelbrecht, J. A. |
| author_sort | Engelbrecht, J. A. |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch University |
| description | Thesis (MEng)--Stellenbosch University, 2021. |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/123607 |
| institution | Stellenbosch University (South Africa) |
| language | en_ZA |
| last_indexed | 2026-06-10T12:42:12.448Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2021 |
| publishDateRange | 2021 |
| publishDateSort | 2021 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/123607 Machine learning for semantic segmentation of biomedical imaging data: ensemble 2D UNets for CT and PET images. Engelbrecht, J. A. Nieuwoudt, M. J. Malherbe, S. T. Stellenbosch University. Faculty of Engineering. Dept. of Mechanical and Mechatronic Engineering. Machine Learning Tuberculosis Semantic Segmentation Computed tomography UCTD PET-CT (Tomography) Thesis (MEng)--Stellenbosch University, 2021. ENGLISH ABSTRACT: In Tuberculosis (TB) treatment response studies the manual delineation of the lung region in Computed Tomography (CT) scans is often required. This is a time consuming process. Manual delineations are required when abnormalities are present in the lung region, such as pronounced pathology caused by TB. Another problem requiring manual correction is misregistration. This is when composite Positron Emission Tomography (PET)/CT scans do not properly align based on anatomical features. Anatomical features in PET scans often extend into corresponding features of CT scans. This is undesirable. We propose a machine learning (ML) solution, based on the U-Net- and Tiramisu architectures for semantic segmentation, to delineate the lung region on CT scans, and correct misregistration on composite PET/CT scans. We create and train 2D implementations of these architectures. To compensate for the lack of spatial context of the 2D models, we combine multiple models trained on the three medical viewing angles to create a more robust ensemble solution, which reduces the amount of trivial false-positives. The CT based models were trained on 72 CT scans, with data augmentations applied. The PET based models were trained on 12 PET scans, also with data augmentation applied. For the CT based models, better results than previously reported were achieved. For the PET based models, improvements might be made with the addition of additional training data in the future. Finally, the CT based ensemble model is made accessible through a user friendly Jupyter notebook created with clinician use in mind. AFRIKAANSE OPSOMMING: In tuberkulose (TB) is behandelingsresponsstudies se handmatige afbakening van die longstreek in CT-skanderings dikwels nodig. Dit is ’n tydrowende pro- ses. Handmatige afbakening is nodig wanneer abnormaliteite in die longgebied voorkom, soos patologie wat deur TB veroorsaak word. Nog ’n probleem wat handmatige regstelling benodig, is misregistrasie. Dit is wanneer saamgestelde PET / CT-skanderings nie behoorlik ooreenstem op grond van anatomiese ken- merke nie. Anatomiese kenmerke in PET-skanderings strek dikwels oor tot in ooreenstemmende kenmerke van CT-skanderings. Dit is ongewens. Ons stel ’n masjienleer (ML) oplossing voor, gebaseer op die U-Net- en Tiramisu argi- tekture, vir semantiese segmentering om die longgebied op CT-skanderings af te baken, en om misregistrasie op saamgestelde PET / CT-skanderings reg te stel. Ons skep 2D-implementasies van hierdie argitekture. Om te voorsien vir die gebrek aan ruimtelike konteks van die 2D-modelle, kombineer ons verskeie 2D-modelle wat op die drie mediese kykhoeke opgelei is, om ’n meer robuuste ensemble-oplossing te skep, wat die eenvoudige valse positiewe voorspellings verminder. Die CT-gebaseerde modelle is opgelei op 72 CT-skanderings, met data-aanvullings toegepas. Die PET modelle is op 12 PET-skanderings opge- lei, ook met die aanvulling van data. Vir die CT modelle is beter resultate behaal as voorheen gerapporteer. Vir die PET-gebaseerde modelle kan ver- beterings aangebring word met die toevoeging van addisionele opleidingsdata. Ten slotte word die CT ensemble-model toeganklik gemaak deur middel van ’n gebruikersvriendelike Jupyter-notaboek wat geskep is met die oog op klinici. Masters 2021-06-23T14:45:15Z 2021-12-22T14:11:57Z 2021-06-23T14:45:15Z 2021-12-22T14:11:57Z 2021-12 Thesis http://hdl.handle.net/10019.1/123607 en_ZA Stellenbosch University 137 pages application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | Machine Learning Tuberculosis Semantic Segmentation Computed tomography UCTD PET-CT (Tomography) Engelbrecht, J. A. Machine learning for semantic segmentation of biomedical imaging data: ensemble 2D UNets for CT and PET images. |
| title | Machine learning for semantic segmentation of biomedical imaging data: ensemble 2D UNets for CT and PET images. |
| title_full | Machine learning for semantic segmentation of biomedical imaging data: ensemble 2D UNets for CT and PET images. |
| title_fullStr | Machine learning for semantic segmentation of biomedical imaging data: ensemble 2D UNets for CT and PET images. |
| title_full_unstemmed | Machine learning for semantic segmentation of biomedical imaging data: ensemble 2D UNets for CT and PET images. |
| title_short | Machine learning for semantic segmentation of biomedical imaging data: ensemble 2D UNets for CT and PET images. |
| title_sort | machine learning for semantic segmentation of biomedical imaging data ensemble 2d unets for ct and pet images |
| topic | Machine Learning Tuberculosis Semantic Segmentation Computed tomography UCTD PET-CT (Tomography) |
| url | http://hdl.handle.net/10019.1/123607 |
| work_keys_str_mv | AT engelbrechtja machinelearningforsemanticsegmentationofbiomedicalimagingdataensemble2dunetsforctandpetimages |