PhD thesis defense to be held on September 14, 2023, at 15:30 (Multimedia Amphitheater, NTUA Central Library)
Picture Credit: Alexia Tzalavra
Thesis title: Development of multiresolution analysis and fuzzy logic techniques to support the computer-aided diagnosis of breast cancer through images of Dynamic Contrast Enhanced- Magnetic Resonance Imaging (DCE-MRI)
Abstract: The present thesis aims at the design, development, and evaluation of machine learning techniques to support the computer diagnosis of breast cancer. Breast cancer is one of the most common causes of death for women wordwide. This fact places the early detection of the disease as one of the major challenges in the field of biomedical engineering. Towards that direction, the Dynamic Contrast Enhaced Magnetic Resonance Imaging (DCE-MRI) is nowadays an area of high research interest focus since it is a diagnostic method with high sensitivity that allows the study of the morphology of the tissue tumors.
Initially, the preprocessing of breast DCE-MRI images is studied focusing on the enhancement of the image quality. As part of the image prerocessing the tumor segmentation from the original images has been studied with the valuable help of a breast radiologist. The goal is the identification of the most important breast anatomical differences as well as the isolation of the regions of interest from the background.
Methodologies of multiscale analysis are further developed to calculate the tumor (segmented regions). The goal is the extraction of texture features which are of major importance for breast cancer detection. The feature selection methods for the discrimination between benign and malignant findings are also studied.
Furthermore, a hybrid classifier combining the features of an Adaptive Neuro-Fuzzy Infererence System (ANFIS) with the Particle Swarm Optimization (PSO) algorithm is developed. The goal is the automatical discrimination between the malignant and benign tumors with the best possible accurance.
A comparative assessment of the developed methodologies for the analysis and the classification of breast DCE-MRI images is also presented. The goal is the evalution of the potential of each of these methods in the classification accuracy for benign and malignant tumors in daily clinical practice. More specifically, the studied methodologies are evaluated in real patient data that were used for research purposes. Breast DCE – MRI images were used from a total of 44 patients (23 with malignant tumors and 23 with benign tumors) collected from the Penn Medicine, Radiology department of the University of Pennsylvania in USA. The proposed multiscale analysis methodologies are assessed in terms of their ability to discriminate benign and malignant tumors. The features extracted by each methodology are fed as input to a series of known classifiers. In terms of classification accuracy, the three-level Stationary Wavelet Transform (SWT) with sym9 as the mother wavelet function outperforms (91% accuracy) the Discrete Wavelet Transform (DWT) when the extracted texture features are fed in a Linear Discriminant analysis (LDA) classifier in a leave-one-out cross validation scheme. In addition, the four-level fast discrete curve transform (FDCT) achieves the maximum classification accuracy (93.18 %) when the extracted features feed the same classifier. Furthermore, the classification accuracy of the developed hybrid classifier is evaluated compared to known classifiers. The investigated classifiers are based on ensembles of neural networks trained by the bagging method, ensembles of feedforward neural networks of different number of hidden neurons and layers, classifiers based on binary logistic regression, Bayesian approach and decision trees. The findings indicate that the proposed hybrid ANFIS-PSO classifier when fed with texture features extracted by the FDCT methodology using four levels of decomposition, outperforms all the investigated breast tumor classification schemes in terms of classification accuracy (94 %), as well as the area under the Receiver Operating Characteristic (ROC) curve.
Supervisor: Professor Konstantina S. Nikita
PhD Student: Alexia Tzalavra