Histopathology involves the examination of tissue sections to identify diseases and represents the gold standard for cancer diagnostics. The manual analysis of tissue samples is a time-consuming task, performed by medical doctors specialized in pathology, called pathologists. However, the agreement between pathologists on diagnoses is still often low, due to several factors (the heterogeneous morphologies of tissue structures, the arbitrary selection of the regions to analyze in detail and the findings evaluation that may be subjective or biased by pathologist experience). Tissue analysis is still usually performed with limited diagnostic assistance in clinical practice, even though digital pathology is becoming more present in clinical routine.
Digital pathology involves acquiring and managing digitized histopathology images, called Whole Slide Images (WSI). WSIs are generally acquired at high-resolution and stored in a multi-scale format, allowing pathologists to visualize different details of the tissue structures, from the lowest to the highest magnification levels. Pathological findings, including observations from WSI analysis, are usually described in a pathology case report, which are semi-structured free-text reports.
Computational pathology is a domain focusing on the development of computer-assisted diagnosis tools to automatically analyze digital pathology images. Convolutional Neural Network have become the state-of-the-art algorithm to solve several computational pathology tasks, such as WSI classification, achieving high accurate performance.
However, many challenges are still open in the computational pathology domain. First, CNNs usually need large datasets for training models that are robust to the high data variability of clinical practice data. Second, fully supervised approaches, reaching the highest performance in several computational pathology tasks, require local annotations that are challenging to obtain in medical contexts as they are time-consuming and require experts performing them. Third, WSIs can be highly heterogeneous in terms of stain and magnification scale due to the lack of standardization in tissue acquisition, leading to low model generalization on data including different acquisition parameters than those included in the data used to train the models. Finally, linking visual and semantic information from reports using machine learning is a challenging and currently not a well-solved task.
This thesis aims to alleviate the limitations related to these challenges, that prevent the application of computational pathology algorithms in clinical practice, with the following objectives: 1) alleviating the need for annotations in computational pathology 2) improving the convolutional neural network capability to generalize on datasets including heterogeneous color variations, 3) combining multiple magnification levels to improve the whole slide image representation learnt by convolutional neural networks 4) empowering the histopathology data representation combining textual and visual information from images and reports.