Machine Learning-based Prediction of Molecular Subtypes of Breast Cancer using DCE MRI - Master thesis - research and results analysis

Aug 3, 2026

Breast cancer is classified into four molecular subtypes using genetic and molecular markers, and this work builds a machine learning framework for their prediction. Radiomics features extracted from dynamic contrast-enhanced MRI (DCE MRI), together with clinical features, form a large dataset containing 4428 radiomics features per patient. The pipeline includes preprocessing, feature extraction, initial and final feature selection, and data cleaning with imputation and Local Outlier Factor (LOF), followed by hyperparameter tuning and robustness analysis. Performance is assessed in 4-label, binary, and 3-label settings, achieving up to 85% F1 for binary tasks and improving overall 4-subtype accuracy by 12%.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages100

Reading Time4 h 12 min

AccessFree to download

Abstract

Introduction

Background and Related Work

Methodology

Data Description

Data Collection

Data Filtration

Cropping the Images

Data Segmentation Using Masks

Feature Extraction

Data Preprocessing

Data Integration

Feature Selection

Classification Using ML Algorithms

Hardware Specifications

Results and Discussion

4-label Classification

Binary Classifications

One Versus the Rest (OvR)

One Versus One (OvO)

Two Versus Two (TvT)

3-label Classifications

3-Label with Elimination of One Label

3-Label with Combining One Label

Conclusion

Principal Findings

Study Limitations

Future Work

Machine Learning-based Prediction of Molecular Subtypes of Breast Cancer using DCE MRI by © Javad Aghadavood Marnani A Thesis submitted to the School of Graduate Studies in partial fulfillment of the requirements for the degree of Master of Science. Supervisor: Dr. Hamid Usefi Co-supervisor: Dr. J. Concepti´on Loredo-Osti Department of Mathematics and Statistics Memorial University of Newfoundland May 2023 St. John’s, Newfoundland and Labrador, Canada Abstract Breast cancer is a prevalent disease that can be classified into four molecular subtypes based on genetic and molecular markers. This study aimed to develop a machine learning-based approach to classify molecular subtypes of breast cancer using radiomics features extracted from dynamic contrast-enhanced magnetic resonance imaging (DCE MRI) . The comprehensive dataset used in this study included 4428 radiomics features per patient, as well as clinical features, making it a valuable resource for future research. Our methodology involved several stages, including image preprocessing, feature extraction, initial and final feature selection, and data cleaning techniques, such as data imputation and Local Outlier Factor (LOF), to ensure the quality of the dataset. We conducted hyperparameter tuning and robustness analysis to optimize the performance of the machine learning algorithms. The results were evaluated in three scenarios: 4-label classification, binary, and 3-label classifications. Our approach achieved up to 85% F1 score in binary classifications and improved the overall accuracy of classifying the four molecular subtypes of breast cancer by 12%, which represents a significant improvement over the original study. These findings suggest that machine learning algorithms can be a powerful tool for improving the diagnosis and treatment of breast cancer, paving the way for personalized medicine approaches. Furthermore, the proposed approach can be applied to other datasets and may be useful in other areas of medical research that rely on radiomics features extracted from medical images. Acknowledgments I would like to express my deepest gratitude to my esteemed supervisors, Dr. Hamid Usefi and Dr. J Concepcion Loredo-Osti, for their invaluable guidance, support, and mentorship during my master’s degree. Their expertise and dedication have been instrumental in shaping my research and academic experience. I extend my profound appreciation to Dr. Alexander Bihlo, Dr. Lourdes PenaCastillo, Dr. Candemir Cigsar, and Dr. Armin Hatefi for their invaluable support and teaching. Furthermore, I would like to thank my family and friends for their unwavering encouragement and support throughout my academic journey. Their continuous support and motivation have been a source of strength and inspiration for me throughout my academic journey. Contents Abstract 1 Acknowledgments 2 List of Figures 6 List of Tables 8 List of Abbreviations 9 1 Introduction 13 2 Background and Related Work 20 3 Methodology 24 3.1 Data Description ................................ 24 3.2 Data Collection ................................. 31 3.3 Data Filtration ................................. 34 3.4 Cropping the Images .............................. 36 3.5 Data Segmentation Using Masks ...................... 38 3.5.1 Image segmentation .......................... 38 3.5.2 Generating Masks Using Thresholding ............... 39 3.6 Feature Extraction ............................... 40 3.7 Data Preprocessing .............................. 45 3.7.1 Data Averaging ............................ 45 3.7.2 Data Splitting ............................. 46 3.7.3 Data Scaling .............................. 46 3.7.4 Data Cleaning ............................. 47 3.8 Data Integration ................................ 49 3.9 Feature Selection ................................ 51 3.9.1 Initial Feature Selections ....................... 51 3.9.1.1 Feature

FAQs

What data and imaging modality are used to predict breast cancer molecular subtypes?

The study uses radiomics features extracted from dynamic contrast-enhanced MRI (DCE MRI) combined with clinical features. Each patient contributes 4428 radiomics features.

How is dataset quality ensured before model training?

The methodology includes preprocessing and data cleaning steps such as data imputation and outlier handling with Local Outlier Factor (LOF). It also performs data scaling and other preprocessing operations.

How is model performance evaluated across different subtype settings?

Results are reported in three scenarios: 4-label classification, binary classification, and 3-label classification. The approach reaches up to 85% F1 score in binary tasks and improves overall 4-subtype accuracy by 12%.

Aug 3, 2026

Upload by Valentina