Data-driven Representation Learning from Histopathology Image Databases to Support Digital Pathology Analysis

Cancer research is a major public health priority in the world due to its high incidence, diversity and mortality. Despite great advances in this area during recent decades, the high incidence and lack of specialists have proven that one of the major challenges is to achieve early diagnosis. Improved early diagnosis, especially in developing countries, plays a crucial role in timely treatment and patient survival. Recent advances in scanner technology for the digitization of pathology slides and the growth of global initiatives to build databases for cancer research have enabled the emergence of digital pathology as a new approach to support pathology workflows. This has led to the development of many computational methods for automatic histopathology image analysis, which in turn has raised new computational challenges due to the high visual variability of histopathology slides, the difficulty in assessing the effectiveness of methods (considering the lack of annotated data from different pathologists and institutions), and the need of interpretable, efficient and feasible methods for practical use. On the other hand, machine learning techniques have focused on exploiting large databases to automatically extract and induce information and knowledge, in the form of patterns and rules, that allow to connect low-level content with its high-level meaning. Several approaches have emerged as opposed to traditional schemes based on handcrafted features for data representation, which nowadays are known as representation learning. The objective of this thesis is the exploration, development and validation of precise, interpretable and efficient computational machine learning methods for automatic representation learning from histopathology image databases to support diagnosis tasks of different types of cancer. The validation of the proposed methods during the thesis development allowed to corroborate their capability in several histopathology image analysis tasks of different types of cancer. These methods achieve good results in terms of accuracy, robustness, reproducibility, interpretability and feasibility suggesting their potential practical application towards translational and personalized medicine.

• Autores: Cruz Roa, Angel Alfonso
• Tipo de recurso: Doctoral thesis
• Fecha de publicación: 2015
• Institución: Universidad Nacional de Colombia
• Repositorio: Universidad Nacional de Colombia
• Idioma: spa
• OAI Identifier: oai:repositorio.unal.edu.co:unal/55193
• Acceso en línea: https://repositorio.unal.edu.co/handle/unal/55193; http://bdigital.unal.edu.co/50510/
• Palabra clave: 61 Ciencias médicas; Medicina / Medicine and health; 62 Ingeniería y operaciones afines / Engineering; Digital pathology; Histopathology image analysis; Representation learning; Deep Learning; Whole slide images; Patología digital; Análisis de imágenes de histopatología; Aprendizaje de la representación; Aprendizaje profundo; Láminas virtuales de patología
• Rights: openAccess
• License: Atribución-NoComercial 4.0 Internacional