Computational methodologies for radiogenomics and digital pathology

The primary objective of this thesis is to showcase computational methodologies for Radiogenomics and Digital Pathology. Radiogenomics seeks to establish connections between a lesion's phenotypic features and its genotypic traits, relying on quantitative insights referred to as radiomics features. Digital Pathology, initially centered on digitizing classical histopathology, has expanded its scope. It now encompasses a broad range of image processing algorithms for analyzing acquired images. Both these disciplines heavily rely on Machine Learning (ML) and Deep Learning (DL) techniques. DL, in particular, has revolutionized medical image analysis, significantly boosting performance in classification, detection, and various medical domains. The thesis focuses on creating accessible, interpretable end-to-end pipelines utilizing ML and DL frameworks, drawing data from both public repositories and local hospitals. In the field of Radiogenomics, the research activities have focused on analyzing lung cancer cases. As a first step, a model that classifies lung adenocarcinomas from other types of lesions using radiomic features extracted from Computed Tomography (CT) images has been developed. Subsequently, a system that classifies the mutational status of two crucial genes in cases of lung adenocarcinomas, KRAS and EGFR, also based on radiomic features extracted from CT images has been built. Predicting the mutational status of these genes is indeed crucial in clinical settings as it enables physicians to tailor a personalized treatment plan for the patient. In the domain of Digital Pathology, the research has concentrated on two studies: the first one assessed the impact of unpaired image-to-image translation (UI2IT) architectures for normalizing hematoxylin and eosin stained images in the classification of histopathological tissue of patients with colorectal cancer. The UI2IT architectures were compared with classical histological image normalization techniques, revealing enhanced classifying performance when images were normalized using a UI2IT model as opposed to classic techniques. The second study involved the development of a system that, starting from Periodic Acid-Schiff Whole Slide Images, segments glomeruli and classifies glomerular lesions according to the Oxford classification in patients with IgA nephropathy. Object detection architectures, particularly Mask R-CNN and Cascade Mask R-CNN, were employed for the segmentation module, while the classification part was achieved through convolutional neural networks. Additionally, the intraclass correlation coefficient was computed for glomerular lesion classification between annotations by an expert pathologist and the classification model results. For each lesion, at least one of the models surpassed the minimum ICC threshold as set by the Oxford classification. The developed tool has been made available on GitHub along with the trained models for use with other images. The thesis's structure is as follows: Chapter 1 introduces the thesis's goals and contributions. Chapter 2 provides a comprehensive overview of the methodologies employed, delving into various ML and DL methodologies and defining Radiomics features. Chapter 3 details the thesis's contributions in Radiogenomics, specifically focusing on lung adenocarcinoma radiomic characterization and the prediction of EGFR and KRAS gene mutational status in lung adenocarcinoma. Chapter 4 outlines the contributions in Digital Pathology, encompassing the role of unpaired image-to-image translation stain color normalization in colorectal cancer histology classification, and the segmentation of glomeruli and Oxford classification of MESC lesions for IgA nephropathy. Lastly, Chapter 5 summarizes the work accomplished in this thesis and offers insights into potential future works.