Dissertation topics suggested by the teacher.
Evaluating Self-Supervised Algorithms for Imbalanced Medical Data
Self-supervised learning is a machine learning technique where a model learns to extract meaningful representations from unlabeled data by creating its own labels or objectives. It enables the model to capture useful information and transfer it to other tasks without relying on explicit human annotations.
This thesis aims to explore the application of self-supervised algorithms in addressing the challenges posed by imbalanced medical data within the field of artificial intelligence for medicine. Imbalanced datasets, where the number of samples in different classes is significantly uneven, are prevalent in medical domains and pose significant obstacles to accurate and reliable AI-based systems. In this study, we will investigate and evaluate the robustness of various self-supervised learning approaches to this issue.
Through rigorous experimentation and evaluation, this thesis aims to provide valuable insights into the strengths and limitations of self-supervised learning in the context of imbalanced medical data. The outcomes of this research will contribute to enhancing the development of robust and accurate AI models for medical applications, ultimately improving diagnosis, treatment, and patient care in the healthcare industry.
The data can be imaging or tabular, e.g., classification of patients with Alzheimer's disease from brain MRI images.
Exploring the role of classic data augmentation in GAN-based conditional synthesis: towards high-quality synthetic medical imaging generation
This thesis aims to investigate and compare the impact of training a Generative Adversarial Network (GAN) on a normal dataset versus a dataset augmented with classical data augmentation techniques. The ultimate aim is to enhance the quality of synthetic images generated through conditional synthesis and potentially realize a high-quality synthetic dataset.
Through a comprehensive evaluation, this thesis explores how applying classic data augmentation techniques, such as flipping, cropping, and other commonly used transformations, can improve the performance and fidelity of GAN-based conditional synthesis. The research will utilize the ADNI (Alzheimer's Disease Neuroimaging Initiative) dataset, which comprises brain MRI scans of individuals diagnosed with Alzheimer's Disease as well as healthy control subjects. By systematically comparing the results obtained from training GANs on both augmented and unaugmented datasets, the study aims to provide insights into the effectiveness of classical data augmentation in synthetic image generation for various medical applications.
The findings of this research will contribute to advancing the field of GAN-based conditional synthesis by identifying the approaches that lead to superior-quality synthetic images. The outcomes will have significant implications for applications such as medical imaging, where access to large and diverse datasets is often limited.
Investigating the impact of synthetic data on CNN classification performance: a comparative study with classic data augmentation techniques
This thesis aims to assess the potential of synthetic data in improving the classification performance of Convolutional Neural Networks (CNNs). The study involves three distinct conditions: training a CNN with a normal dataset, training a CNN with a dataset augmented using classical data augmentation techniques, and training a CNN with a synthetic dataset generated by a conditional GAN capable of performing conditional image synthesis.
The dataset used in this research will be the ADNI (Alzheimer's Disease Neuroimaging Initiative) dataset, consisting of brain MRI scans of Alzheimer's Disease (AD) patients and healthy control (HC) subjects. The synthetic dataset will be created using an existing conditional GAN, specifically designed by our research group for the synthesis of brain MRI scans of AD and HC. Through a comparative analysis, this study seeks to determine whether training a CNN with synthetic data can yield improved classification performance compared to training with traditional datasets and augmented datasets.
The findings of this study will provide valuable insights into the potential benefits of synthetic data and its ability to enhance CNN-based classification tasks. This research has significant implications for various domains, such as medical imaging and computer vision, where limited access to labeled datasets can be a bottleneck.
