Deep learning for processing histopathology images

Read full text Summarise with VIRAT
Authors: Andrew Moyes
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Histopathology is the study and diagnosis of disease via tissue microscopy and it is currently the ‘gold-standard‘ in formally diagnosing many types of disease including cancers. Due to increasing workloads on pathologists, there is a growing need for automated image analysis pipelines that are able to filter out obviously benign samples. However, these algorithms are sensitive to factors of variation such as the staining and scanning conditions with which a tissue specimen is processed that differ significantly across institutions. In this thesis, novel machine learning algorithms are developed using state-of-the-art deep learning techniques that enable the creation of automated histopathology image analysis pipelines that are more robust to frequently occurring artifacts and variation in appearance. A novel neural network architecture called the Dual-Channel Auto-Encoder (DCAE) is devised that learns representations of histopathology images that are invariant to the effects of differing digital slide scanners whilst holding enough discriminative power to delineate various anatomical structures within tissue. This method achieves a 50% improvement in SSIM score on tissue masks derived from the DCAE features compared to related methods. Next, a novel unsupervised approach to detect and remove artifacts using techniques from generative adversarial networks is presented. This approach allows artifacts to be removed prior to the application of stain separation which is demonstrated experimentally to improve the stain colour estimation performance. Finally, a more sophisticated model of stain separation called Tissue-Dependent Stain Separation (TDSS) is developed which incorporates contextual information such as texture and colour to make more informed estimates of stain colour conditioned on tissue sub-types and thus producing better separations of stains compared to state-of-the-art methods. The TDSS model achieves an 8.81% reduction in mean-squared error in a stain separation task compared to existing state-of-the-art-methods. These contributions represent a significant step forward towards robust, fully automated histopathology image analysis pipelines.

Fields of Study & Subjects

AI in cancer detection Cell Image Analysis Techniques Digital Imaging for Blood Diseases

Publication details

Journal / source
Research Portal (Queen's University Belfast)
Publication date
1 October 2026
Language
EN
Document type
Thesis