Abstract

Recent advancements in signal processing and machine learning coupled with developments of electronic medical record keeping in hospitals have resulted in a surge of significant interest in ``radiomics". Radiomics is an emerging and relatively new research field, which refers to semi-quantitative and/or quantitative features extracted from medical images with the goal of developing predictive and/or prognostic models. Radiomics is expected to become a critical component for integration of image-derived information for personalized treatment in the near future. The conventional radiomics workflow is, typically, based on extracting pre-designed features (also referred to as hand-crafted or engineered features) from a segmented region of interest. Clinical application of hand-crafted radiomics is, however, limited by the fact that features are pre-defined and extracted without taking the desired outcome into account. The aforementioned drawback has motivated trends towards development of deep learning-based radiomics (also referred to as discovery radiomics). Discovery radiomics has the advantage of learning the desired features on its own in an end-to-end fashion. Discovery radiomics has several applications in disease prediction/ diagnosis. Through this Ph.D. thesis, we develop deep learning-based architectures to address the following critical challenges identified within the radiomics domain. First, we cover the tumor type classification problem, which is of high importance for treatment selection. We address this problem, by designing a Capsule network-based architecture that has several advantages over existing solutions such as eliminating the need for access to a huge amount of training data, and its capability to learn input transformations on its own. We apply different modifications to the Capsule network architecture to make it more suitable for radiomics. At one hand, we equip the proposed architecture with access to the tumor boundary box, and on the other hand, a multi-scale Capsule network architecture is designed. Furthermore, capitalizing on the advantages of ensemble learning paradigms, we design a boosting and also a mixture of experts capsule network. A Bayesian capsule network is also developed to capture the uncertainty of the tumor classification. Beside knowing the tumor type (through classification), predicting the patient's response to treatment plays an important role in treatment design. Predicting patient's response, including survival and tumor recurrence, is another goal of this thesis, which we address by designing a deep learning-based model that takes not only the medical images, but also different clinical factors (such as age and gender) as inputs. Finally, COVID-19 diagnosis, another challenging and crucial problem within the radiomics domain, is dealt with using both X-ray and Computed Tomography (CT) images (in particular low-dose ones), where two in-house datasets are collected for the latter and different capsule network-based models are developed for COVID-19 diagnosis.