Localisation and symmetry in computational pathology

Preview

PDF WRAP_Theses_Graham_2020.pdf - Submitted Version - Requires a PDF viewer. Download (104MB) | Preview

Request Changes to record.

Abstract

Conventional assessment of Haematoxylin and Eosin (H&E) stained tissue slides is performed via visual examination under the microscope by a pathologist and often serves as the gold standard in cancer diagnosis. Standard diagnostic practice requires pathologists to follow a descriptive set of guidelines and is, therefore, prone to suffer from inter-observer variability due to differences in interpretation of histological patterns. Furthermore, each tissue slide may contain tens of thousands of cells and, therefore, accurate quantification and morphological analysis of the tissue in the entire slide is not feasible. Recently, there has been a growing trend towards a digital pathology workflow, where tissue slides are digitised with a high-resolution scanner to obtain Whole-Slide Images (WSIs). This enables the development of automatic tools that can objectively analyse and quantify the vast amount of pixel information contained in multi-gigapixel WSIs.

In this thesis, we initially introduce the challenge of analysing large-scale WSIs for histology image analysis by presenting a preliminary WSI classification framework. Here, we predict the diagnosis of a slide by: (i) dividing the WSI into small image regions (patches), (ii) making predictions independently on each patch and then (iii) predicting the overall slide diagnosis by aggregating patch-level results.

In the remainder of the thesis, we focus on developing automated methods that localise objects and structures of interest in the tissue and that leverage the presence of rotational symmetry in histology images. Localisation of nuclei and other components, such as glands, allows further exploration of digital biomarkers and serves as a fundamental pre-requisite for downstream analysis. On the other hand, exploitation of rotational symmetry for histology image analysis enables models to be tailored to the specific geometry of microscopy images, where there exists no underlying global orientation.

In this regard, we present the first single convolutional neural network (CNN) for simultaneous segmentation and classification of nuclei. The CNN uses the concept of horizontal and vertical maps to separate clustered nuclei and utilises a devoted upsampling branch to accurately perform nuclear classification. We then propose a novel CNN for gland segmentation that counters the loss of information caused by max-pooling by reintroducing the original image at multiple points within the network. To enable localisation of glands with varying size, we additionally incorporate atrous spatial pyramid pooling.

To leverage the prior knowledge that histology images are symmetric under rotation, it is desirable for CNNs to be rotation-equivariant. This guarantees that features transform as expected with rotation of the input. In this thesis, we perform the first thorough analysis of various rotation-equivariant models for histology image analysis. We then develop a CNN for simultaneous segmentation of glands and lumen that achieves rotation-equivariance by using group-convolutions with multiple rotated copies of each filter. Finally, we propose a general CNN for histology image analysis that employs the concept of group-convolution and defines filters as a linear combination of steerable basis filters. This enables exact rotation and decreases the number of trainable parameters compared to standard filters.

Request changes or add full text files to a record

Repository staff actions (login required)

View Item