Deep Learning for Medical Imaging Applications

Over the past decade, deep learning algorithms have been proved to excel at various medical imaging tasks ranging from disease detection to progression prediction and segmentation. One of the primary goals of medical imaging is to improve clinical translation enabling generalization across different real-world clinical settings and patient populations. However, there are some major challenges in real-world data applications pulling down the progress including lack of datasets reflecting real-world clinical data, clinical interpretability, generalization to real-world clinical settings, and learning from extremely small datasets that is inevitable in many medical. My thesis focuses on addressing these challenges in the context of ophthalmic imaging applications in four settings. My first work focuses on developing a new medical a new ophthalmic imaging dataset where the data is collected from real-world clinical data to provide an infrastructure for validation studies and improving the clinical translation. In my second work, I propose a multi-task learning model that achieves clinical interpretability for a real-world clinical application by exploiting the complementary information across related tasks. My third work presents a feature extraction network for ophthalmic imaging applications by employing a self-supervised learning framework. In this work, we also assess deep learning models and their generalization capacity in coping with real-world data versus standardized data. This works also studies the effectiveness of self-supervised contrastive learning in learning more generalizable features. In my last work, I present a model that harnesses the power of segmentation to learn from small datasets allowing us to perform classification on extremely small datasets.