Biomedical image analysis of brain tumours through the use of artificial intelligence
| dc.contributor.advisor | Muller, C. J. B. | en_ZA |
| dc.contributor.author | Di Santolo, Claudia | en_ZA |
| dc.contributor.other | Stellenbosch University. Faculty of Economic and Management Sciences. Dept. of Statistics and Actuarial Science. | en_ZA |
| dc.date.accessioned | 2022-03-09T07:00:48Z | |
| dc.date.accessioned | 2022-04-29T09:25:03Z | |
| dc.date.available | 2022-03-09T07:00:48Z | |
| dc.date.available | 2022-04-29T09:25:03Z | |
| dc.date.issued | 2022-04 | |
| dc.description | Thesis (MCom)--Stellenbosch University, 2022. | en_ZA |
| dc.description.abstract | ENGLISH SUMMARY: Cancer is one of the leading causes of morbidity and mortality on a global scale. More specifically, cancer of the brain, which is one of the rarest forms. One of the major challenges is that of timely diagnoses. In the ongoing fight against cancer early and accurate detection in combination with effective treatment strategy planning remains one of the best tools for improved patient outcomes and success. Emphasis has been placed on the identification and classification of brain lesions in patients - that is, either the absence or presence of brain tumours. In the case of malignant brain tumours it is critical to classify patients into either high-grade or low-grade brain lesion groups: different gradings of brain tumours have different prognoses, thus different survival rates. The growth in the availability and accessibility of big data due to digitisation has led individuals in the area of bioinformatics in both academia and industry to apply and evaluate artificial intelligence techniques. However, one of the most important challenges, not only in the field of bioinformatics but also in other realms, is transforming the raw data into valuable insights and knowledge. In this research thesis artificial intelligence techniques that can detect vital and fundamental underlying patterns in the data are reviewed. The models may provide significant predictive performance to assist with decision making. Much artificial intelligence has been applied to brain tumour classification and segmentation in the research literature. However, in this study the theoretical background of two more traditional machine learning methods, namely 𝑘-nearest neighbours and support vector machines, is discussed. In recent years, deep learning (artificial neural networks) has gained prominence due to its ability to handle copious amounts of data. The specialised version of the artificial neural network that is reviewed is convolutional neural networks. The rationale behind this particular technique is that it is applied to visual imagery. In addition to making use of the convolutional neural network architecture, the study reviews the training of neural networks that involves the use of optimisation techniques, considered to be one of the most difficult parts. Utilising only one learning algorithm (optimisation technique) in the architecture of convolutional neural network models for classification tasks may be regarded as insufficient unless there is strong support in the design of the analysis for using a particular technique. Nine state-of-the-art optimisation techniques formed part of a comparative study to determine if there was any improvement in the classification and segmentation of high-grade or low-grade brain tumours. These machine learning and deep learning techniques have proved to be successful in image classification and - more relevant to this research – brain tumours. To supplement the theoretical knowledge, these artificial intelligence methodologies (models) are applied through the exploration of magnetic resonance imaging scans of brain lesions. | en_ZA |
| dc.description.abstract | AFRIKAANSE OPSOMMING: Kanker is wêreldwyd een van die hoofoorsake van morbiditeit en sterftes; veral breinkanker, wat een van die mees seldsame soorte is. Een van die groot uitdagings is om dit betyds te diagnoseer. In die voortgesette stryd teen kanker is vroeë en akkurate opsporing, in kombinasie met doeltreffende beplanning van die behandelingstrategie, een van die beste hulpmiddels vir verbeterde pasiëntuitkomste en sukses. Klem word geplaas op die identifikasie en klassifikasie van breinletsels in pasiënte – dit wil sê, die teenwoordigheid of afwesigheid van breingewasse. In die geval van kwaadaardige breingewasse is dit noodsaaklik om pasiënte in groepe as hetsy hoëgraad- of laegraadbreingewasse te klassifiseer: verskillende graderings van breingewasse het verskillende prognoses, en dus verskillende oorlewingskoerse. Die toename in die beskikbaarheid en toeganklikheid van groot data danksy digitalisering, het daartoe gelei dat individue op die gebied van bio-informatika in die akademie en die bedryf begin het om kunsmatige-intelligensie-tegnieke toe te pas en te evalueer. Een van die belangrikste uitdagings, nie slegs op die gebied van bio-informatika nie, maar ook op ander terreine, is egter die omskakeling van rou data na waardevolle insigte en kennis. Hierdie navorsingstesis hersien die kunsmatige-intelligensie-tegnieke wat lewensbelangrike en grondliggende onderliggende patrone in die data kan opspoor. Die modelle kan beduidende voorspellende prestasie bied om met besluitneming te help. Die navorsingsliteratuur dek heelwat toepassings van kunsmatige intelligensie op breingewasklassifikasie en -segmentasie. In hierdie studie word die teoretiese agtergrond van meer tradisionele masjienleermetodes, naamlik die 𝑘-naaste-bure-algoritme (𝑘-nearest neighbour algorithm) en steunvektormasjiene, bespreek. Diep leer (kunsmatige neurale netwerke) het onlangs op die voorgrond getree weens die vermoë daarvan om groot hoeveelhede data te kan hanteer. Die gespesialiseerde weergawe van die kunsmatige neurale netwerk wat hersien word, is konvolusionele neurale netwerkargitektuur. Die rasionaal vir hierdie spesifieke tegniek is dat dit op visuele beelde toegepas word. Buiten dat dit van konvolusionele neurale netwerkargitektuur gebruik maak, hersien die studie ook die afrigting van neurale netwerke met behulp van optimaliseringstegnieke, wat as een van die moeilikste dele beskou word. Die aanwending van slegs een leeralgoritme (optimaliseringstegniek) in die argitektuur van konvolusionele neurale netwerkmodelle vir klassifikasietake, kan as onvoldoende beskou word, tensy daar sterk steun vir die gebruik van ʼn spesifieke tegniek in die ontwerp van die ontleding is. Nege van die jongste optimaliseringstegnieke was deel van ʼn vergelykende studie om vas te stel of daar enige verbetering in die klassifikasie en segmentasie van hoëgraad- en laegraadbreingewasse was. Hierdie masjienleer- en diep-leertegnieke was suksesvol met beeldklassifikasie en – meer relevant vir hierdie navorsing – breingewasklassifikasie. Ter aanvulling van die teoretiese kennis, word hierdie kunsmatige-intelligensie-metodologieë (-modelle) deur die verkenning van magnetiese resonansbeelding van breingewasse toegepas. | af_ZA |
| dc.description.version | Masters | |
| dc.format.extent | xvi, 235 pages : illustrations, includes annexures | |
| dc.identifier.uri | http://hdl.handle.net/10019.1/124661 | |
| dc.language.iso | en_ZA | en_ZA |
| dc.publisher | Stellenbosch : Stellenbosch University | |
| dc.rights.holder | Stellenbosch University | |
| dc.subject | Imaging systems in medicine -- South Africa | en_ZA |
| dc.subject | Deep learning (Machine learning) -- South Africa | en_ZA |
| dc.subject | Cancer -- Prognosis -- South Africa | en_ZA |
| dc.subject | Neural networks (Computer science) -- South Africa | en_ZA |
| dc.subject | Artificial intelligence -- Medical applications -- South Africa | en_ZA |
| dc.subject | UCTD | |
| dc.title | Biomedical image analysis of brain tumours through the use of artificial intelligence | en_ZA |
| dc.type | Thesis | en_ZA |
Files
Original bundle
1 - 1 of 1
Loading...
- Name:
- santolo_biomedical_2022.pdf
- Size:
- 3.63 MB
- Format:
- Adobe Portable Document Format
- Description: