Abstract

Early diagnosis and prognosis of life-threatening diseases such as the novel coronavirus infection (COVID-19) and Lung Cancer (LC), involves tackling critical challenges including but not limited to their undisclosed characteristics, non-stationary nature, significant inter-disease similarities, and intra-disease variations. In particular, within the context of a highly contagious disease such as COVID-19, early and reliable diagnosis is of significant importance. On the other hand, when it comes to diagnosis and prognosis of LC, an accurate prediction of the disease invasiveness becomes of primary importance. Recent advancements of Artificial Intelligence (AI) and Deep Learning (DL)-based architectures have resulted in a surge of interest in the utilization of medical images to develop decision support and stand-alone models to address the aforementioned challenges. In this context, the focus of the thesis is on the utilization of volumetric chest CT images to develop robust and fully automated diagnostic frameworks for COVID-19 diagnosis and LC invasiveness prediction. In particular, Capsule Network (CapsNet) and Transformer-based architectures are developed to expand the application of AI in this domain. More specifically, first, CT-CAPS and COVID-FACT frameworks are proposed to analyze CT images, identify slices demonstrating infection, and perform patient-level classification of COVID-19. The proposed frameworks are developed based on the CapsNet architecture, which unlike the widely-used Convolutional Neural Networks (CNNs), is capable of capturing spatial relations among instances in an image and being trained on small datasets. These characteristics are of utmost importance when analyzing a newly emerged disease with specific spatial patterns in its images. Furthermore, following the recent and ever-increasing interest in using Low-Dose and Ultra-Low-Dose CT scans (LDCT and ULDCT) for COVID-19 screening, the WSO-CAPS framework is proposed to enhance performance of the proposed models to deal with noisy and low-quality CT scans. In addition, given that CT scans acquired from multiple centers and cohorts mainly show different qualities and characteristics, which negatively affect the generalizability of DL-based models, a unique multi-center dataset of CT scans, referred to as the “SPGC-COVID Dataset”, is constructed, which incorporates CT scans of COVID-19, Community Acquired Pneumonia (CAP), and normal cases, obtained using standard and low-dose imaging protocols. An enhancement approach is then proposed to boost the performance of the developed classification frameworks when being tested on varied CT scans in the SPGC-COVID dataset. With respect to the second objective of this thesis (i.e., Lung Cancer invasiveness prediction), the CAE-Transformer framework is proposed, which utilizes image-driven features to predict the invasiveness of Lung Adenocarcinomas (LUACs) from non-thin 3D CT scans. The proposed framework introduces a new viewpoint in CT scan analysis, which relies on the sequential nature of the volumetric CT scans. More specifically, the CAE-Transformer adopts the transformer architecture, which was initially designed for sequential data, to capture inter-slice dependencies in an efficient and non-complex fashion.