Post-graduate theses
Search command : Author="Καραδήμας"
And Author="Ευάγγελος"
Current Record: 16 of 6695
|
Identifier |
000465724 |
Title |
A radiomics analysis pipeline in prostate cancer imaging |
Alternative Title |
Μια διαδικασία ανάλυσης ακτινονομικών δεδομένων στην απεικόνιση του καρκίνου του προστάτη |
Author
|
Ηλία, Ευαγγελία
|
Thesis advisor
|
Μαριάς, Κωνσταντίνος
|
Abstract |
This research aims to answer the question: what is the best preprocessing pipeline for
radiomics analysis of prostate cancer MRI images and how can that benefit the performance
of a radiomics model. To achieve this ~80 papers were compared on their respective
preprocessing pipeline, study design , results and limitations. Through this process a pipeline
proposal was formed which included bias field correction, normalization and resampling.
This pipeline was then tested on the ProstateX dataset which included MRI scans from 66
patients, prostate segmentations and True/False labels for clinical significance. Following
preprocessing the radiomics features were extracted and utilized as the input for model
building. After comparing different classifiers, Logistic Regression was selected as a stand
out. Hyperparameter tuning was used in order to find the best parameters for the model by
utilizing 5-fold repeated stratified cross-validation. The dataset was used to create two
models. The first was divided into a train (70%) set and a test(30%) set. The training set was
used for the tuning and the training whereas the test set was used only at the end for
evaluation of the final model. The model achieved an accuracy of 80.4% on the training, 75%
on the test set and an AUC of 0.867 on the training and 0.791 on the test set. To reduce
overdiagnosing it is important to focus on the precision metrics too. On the training set an
76.9% precision was achieved compared to a 60% on the test set. This difference means that
although on the training set the model was fairly good at avoiding false positives, its
performance on the test set was lacking a bit in comparison. The second model divided the
dataset into 80% training and 20% hold-out set. The training set was utilized the exact same
way as before while the hold-out set was kept aside and used to evaluate the performance
of the trained model. This model did perform worse due to the “unseen data” aspect.
Precision in the training set was 80% but dropped to 50% on the hold-out and AUC went
92,2% to 68,9%. This finding demonstrates clearly the necessity of creating larger publicly
accessible datasets in order to create more reliable models that may eventually be
implemented in clinical settings. Conclusively, this research was successful in proposing an
effective preprocessing pipeline that achieved notable performance results on the final
radiomics model test set but did not do as good on unseen data. Still, this work represents a
significant step forward and may pave the way for more studies and future clinical
applications.
|
Language |
English, Greek |
Subject |
Ακτινονομικά δεδομένα |
Issue date |
2024-07-26 |
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/0/3/c/metadata-dlib-1719475832-668318-28339.tkl
|
Views |
485 |