Post-graduate theses
Search command : Author="Μπριασούλης"
And Author="Γεώργιος"
Current Record: 21 of 1204
|
Identifier |
000457825 |
Title |
Developing a comprehensive radiomic analysis workflow for the detection of prostate cancer aggressiveness on T2weighted MR prostate data |
Alternative Title |
Ανάπτυξη μοντέλων ραδιομικής ανάλυσης για την ανίχνευση της επιθετικότητας του καρκίνου του προστάτη σε Τ2w εικόνες προστάτη μαγνητικής τομογραφίας |
Author
|
Ζαφείρης, Στυλιανός
|
Thesis advisor
|
Μαριάς, Κωνσταντίνος
Μανίκης, Γεώργιος
Ζερβάκης, Μιχάλης
|
Abstract |
Prostate cancer (PCa) is the second most common cancer diagnosed in male population
worldwide, affecting 1.4 million men annually. Early assessment of the malignancy is crucial
for treatment planning and extending patients’ life expectancy. Imaging modalities such as
Magnetic Resonance Imaging (MRI) are used for the non-invasive classification of patients in
order to prevent overtreating indolent malignancies and undertreating those who warrant
immediate treatment. The field of radiomics offers a large quantity of imaging features that
describe the cancer phenotype and can be used in training machine learning (ML) models for
predicting cancer aggressiveness. Effective model training necessitates feature selection,
decreasing the high dimensionality and ensuring the inclusion of pertinent and non-
redundant features. The objective of this study is to investigate the most commonly used
feature selection methods and classifiers in order to predict the tumor’s aggressiveness and
analyze how various image preprocessing techniques affect the performance of the models.
A publicly available multivendor dataset consisting of 225 samples with clinically significant
PCa (csPCa) from 220 patients was used for the analysis. Samples were split in two cohorts
based on ISUP score provided by clinicians. The first cohort (n = 135) contains samples with
an assigned ISUP score equal to 2 (low aggressiveness csPCa) and the second cohort (n = 90)
comprise samples with an assigned ISUP score of 3, 4 and 5 (high aggressiveness csPCa).
Samples with ISUP score equal to 2 tend to have cancer cells that grow slowly, as opposed to
the moderate and quick growth of cancer cells in samples of the second cohort. Thus, early
detection of the tumor grade could prevent an unnecessary intervention or accelerate biopsy.
A comprehensive search for the optimal pipeline was conducted for classifying the
aggressiveness of csPCa. Intensity normalization methods and the N4 bias field correction
method were used to investigate whether these preprocessing steps affect the performance
of the models. For the original and each pre-processed dataset, a cross-combination strategy
leveraging 6 classifiers and 13 feature selection methods was used for determining an optimal
pipeline that reduces overfitting and best determines the tumor grade. Furthermore, hybrid
feature selection methods were also investigated, using the optimal parameter set extracted
from the pipeline. Methods investigated in this study demonstrated a balanced accuracy of
70% in determining the tumor’s aggressiveness, providing promising results in early detection
of aggressiveness of csPCa.
|
Language |
English |
Subject |
Dynamic pipeline optimization |
|
Feature selection |
|
Image preprocessing |
|
Ανάλυση με χρήση ραδιομικών χαρακτηριστικών |
|
Δυναμική βελτιστοποίηση μοντέλων μηχανικής μάθησης |
|
Επεξεργασία εικόνας |
Issue date |
2023-07-28 |
Collection
|
School/Department--School of Medicine--Department of Medicine--Post-graduate theses
|
|
Type of Work--Post-graduate theses
|
Permanent Link |
https://elocus.lib.uoc.gr//dlib/5/3/1/metadata-dlib-1693642352-754369-6727.tkl
|
Views |
384 |
Digital Documents
|
|
No permission to view document.
It won't be available until: 2024-07-28
|