Novel context-based segmentation algorithms for intelligent microscopy

The aim of this proposal is to develop new computer methods to analyse, quantify and understand the information contained in images of cells and tissues obtained using digital microscopy. This will have wide applications in many areas of biomedicine, especially in histopathology, where it can be used to diagnose cancer, predict the potential behaviour of malignant disease, and implement the most suitable form of treatment.
Currently, decisions about histopathological diagnosis rely to a great extent on expert observers and their experience, but this has the disadvantage that the inevitable element of subjectivity in visual observation makes it difficult to reach quantitative or reproducible judgements.
In this project we will design 'context-based' imaging programs to help advance automated analysis and diagnosis of cancer. By 'context-based' here we understand the use of data constructs that (1) allow the structure and relations of cells and tissues in biopsy samples to be modelled in a way that enables computer programs to subsequently 'reason' about the image contents, and (2) allow methods of data extraction that are both quantitative and reproducible.
This will be achieved by using a spatial logic called Discrete Mereotopology (DM). We already have proof-of-concept software and a peer-reviewed preliminary publication which demonstrate the efficacy of this approach for encoding and querying relations between the biologically-relevant entities (e.g. cell nucleus, tissue layers, staining patterns) and the image segments detected to correspond to them, thereby enabling histologically relevant models (e.g. cells, tissue types and voids) to be formulated that can be operated on at a level that has not been possible before. This is a departure from traditional pixel-based routines, leading to region-based algorithms that are histologically relevant to the range of images and structures that are expected in histological imaging.
This logic-based approach to image analysis brings several advantages. Firstly, it provides a robust, rigorous mathematical foundation for software development. Secondly, it allows histological images to be systematically interpreted in terms of histologically-relevant entities, not just pixels. Thirdly, it enables symbolic and automated reasoning programs to be used alongside numerical methods.
We will also incorporate immunohistochemical markers to our histologically relevant models. A key innovation in the proposal is the use of DM to explicitly model the histological localisation of molecular markers, which has never been done before. This will allow a rational evaluation of immunohistochemical patterns and subsequent development of histological predictor markers associated with tumour behaviour and outcome, i.e., what patterns are observed in what kinds of neoplasm, where in tissues the markers are expressed, and how characteristic the patterns are to recorded case outcomes.
Finally we will develop markers of data quality to label structures according to how well they approach the expected models they represent. In this way indications of 'confidence in the results' can be associated to the models found in microscopy images. So far this feature is not provided by any form of histological imaging.
We believe that the approach outlined here is both translational and invaluable in most biological areas using microscopy, where quantitative results are required to make sound evidence-based decisions.

This is a cross-disciplinary translational project where the immediate beneficiaries are cancer patients and the healthcare profession. The overall aim of this research is to address clinical needs in the performance of imaging programs for understanding biological data from microscopy images. While this proposed project specifically targets the diagnosis and prognosis of oropharyngeal carcinomas, the methods adopted are by no means restricted to this type of cancer, or even cancer in general, but will be applicable to most biological areas dealing with interpretation of microscopy images of cells and tissues. Deliverables from this project will make an impact by enabling imaging technologies to elevate the interpretation of biological data (histological imagery) to levels that so far have been difficult to establish.

For the healthcare profession this extends to impact arising from new software tools that enable reliable data extraction and mechanical reasoning on microscopy image contents in ways that have not been possible before. These will provide the means to implement high-throughput systems to analyse more data than is currently feasible, to develop new histological diagnostic standards based on quantifiable evidence-based research, and to create better diagnostic and prognostic models of disease.

The expectations of impact in society are high.
1) Faster and more reliable diagnosis of cancers, resulting in quicker turnaround, with fewer delays in instituting a definitive therapy or adjuvant therapy post-operatively. As outcomes have been shown to be associated with time delay to therapy, faster diagnosis and quicker treatment times are likely to result in better survival rates.
2) Studies have shown that the period of awaiting a definitive diagnosis is the time of highest stress for patients. Faster turnaround in analysing samples will help to ameliorate this.
3) Higher reliability of tests will result in fewer incorrect diagnoses, with less harm to patients and lower health, social and economic costs of implementing incorrect treatments.
4) There is a shortage of histopathologists in the UK. By enabling their work to be more efficient through automation, this will result in a decrease in the delays caused by overstretched resources, and consequently in lower costs, that would translate into efficient use of resources for the NHS.

In terms of benefits to academia and commercial research, the work will advance context-based imaging algorithms that are sensitive to the spatial localisation of features in tissues as it will make it possible to pinpoint not only what cell and tissue patterns are detected, but also where within the samples those features of interest are localised. The research will also provide key insights to the qualitative expression of immunohistochemical markers in images to enable those markers to be related to tissue biological behaviour and therefore to the prognosis of tumours.