Representation learning for histopathology image analysis

Abstract. Nowadays, automatic methods for image representation and analysis have been successfully applied in several medical imaging problems leading to the emergence of novel research areas like digital pathology and bioimage informatics. The main challenge of these methods is to deal with the high visual variability of biological structures present in the images, which increases the semantic gap between their visual appearance and their high level meaning. Particularly, the visual variability in histopathology images is also related to the noise added by acquisition stages such as magnification, sectioning and staining, among others. Many efforts have focused on the careful selection of the image representations to capture such variability. This approach requires expert knowledge as well as hand-engineered design to build good feature detectors that represent the relevant visual information. Current approaches in classical computer vision tasks have replaced such design by the inclusion of the image representation as a new learning stage called representation learning. This paradigm has outperformed the state-of-the-art results in many pattern recognition tasks like speech recognition, object detection, and image scene classification. The aim of this research was to explore and define a learning-based histopathology image representation strategy with interpretative capabilities. The main contribution was a novel approach to learn the image representation for cancer detection. The proposed approach learns the representation directly from a Basal-cell carcinoma image collection in an unsupervised way and was extended to extract more complex features from low-level representations. Additionally, this research proposed the digital staining module, a complementary interpretability stage to support diagnosis through a visual identification of discriminant and semantic features. Experimental results showed a performance of 92% in F-Score, improving the state-of-the-art representation by 7%. This research concluded that representation learning improves the feature detectors generalization as well as the performance for the basal cell carcinoma detection task. As additional contributions, a bag of features image representation was extended and evaluated for Alzheimer detection, obtaining 95% in terms of equal error classification rate. Also, a novel perspective to learn morphometric measures in cervical cells based on bag of features was presented and evaluated obtaining promising results to predict nuclei and cytoplasm areas.

• Autores: Arevalo Ovalle, John Edilson
• Fecha de publicación: 2013
• Institución: Universidad Nacional de Colombia
• Repositorio: Universidad Nacional de Colombia
• Idioma: spa
• OAI Identifier: oai:repositorio.unal.edu.co:unal/49577
• Acceso en línea: https://repositorio.unal.edu.co/handle/unal/49577; http://bdigital.unal.edu.co/43047/
• Palabra clave: 0 Generalidades / Computer science, information and general works; 61 Ciencias médicas; Medicina / Medicine and health; 62 Ingeniería y operaciones afines / Engineering; Histopathology; Image representation; Interpretability; Feature learning; Digital pathology
• Rights: openAccess
• License: Atribución-NoComercial 4.0 Internacional