AI at the forefront of the eye

Summary

The cornea is the only tissue where sensory nerves and immune cells can be non-invasively imaged in humans, using in vivo confocal microscopy. Relevant to both ophthalmology and neurology, corneal confocal microscopy can identify and monitor the immunological and neural effects of ocular and systemic disease. However, its use in research and clinical settings is limited by lack of reliable, time-efficient methods to derive quantitative data from acquired images. A typical confocal imaging session will yield a few hundred images per patient, varying in quality (focusing consistency, field-of-view, degree of artefact). Reliable quantification of cellular features requires selection of a high-quality non-overlapping image set.

Currently, a suitable image analysis set is defined manually, and is time-intensive, poorly repeatable and vulnerable to bias. Quantitative analyses of immune cell morphology and density are also undertaken manually. These factors impede broader implementation of this technology, which has vast capacity to inform diagnostic and treatment decisions. There is a need for reliable, time-efficient, standardised methods for optimised image selection and analysis.

Aim

Our project will advance data science by preparing a high-quality data set for analysis and experimenting with methods supporting the first multi-stage deep learning framework to facilitate quality assessment and streamline analysis of human corneal confocal microscopy images. Whilst application of AI methods to retinal images is established, these approaches have not been leveraged to improve the rigor and efficiency of corneal anatomical imaging.

Objectives

To develop a novel digital platform for robust, automated post-acquisition image selection and immune cell analysis of human corneal confocal images. We will apply machine learning methods to develop and validate a:

1) Triaging tool to automatically identify and select a limited number of high-quality, non-overlapping images for subsequent analysis; this will minimise operator bias and improve analysis efficiency.

2) Model for automated detection and characterisation of immune cells; this will enable rapid and reliable evaluation of corneal inflammation, to advance clinical care and the investigative capacity of the technology for research.

Methodologies

Aim 1: we will use a large set of corneal confocal images for this project. The image sets will comprise a: (i) raw image set; and (ii) quality-refined image set. These image sets will be used to train a machine learning model to identify high-quality images. The triage tool will then automatically select a random subset of non-overlapping images for further analysis.

Aim 2: a trained researcher will manually annotate images in the final analysis sets to: (i) identify immune cells (hyperfluorescent dendritiform structures); and (ii) outline cell borders (to quantify parameters such as field area, perimeter and circularity) in previously-selected, high-quality corneal images. Manual annotations will serve as ground-truth for training and validation of a UNet-like convolutional neural network. Detected and segmented immune cells will be automatically characterised to deliver standardised, objective, and rapid evaluation of ocular surface inflammation.

Progress to date

Aim 1: The development image sets have been curated, and a research assistant has been appointed to develop the machine learning model for identification of high-quality images.

Aim 2: Manual cell annotations have been performed by a trained researcher, and inter-rater reliability estimates are being performed.

Researchers

Chief Investigator

• 
  A/Prof Laura Downie, Associate Professor and Dame Kate Campbell Fellow
  Department of Optometry And Vision Sciences
  Faculty of Medicine, Dentistry and Health Sciences

Co-Investigators

• 
  Prof Karin Verspoor, Professor (Honorary)
  Formerly:

  UoM, School of Computing and Information Systems
  Faculty of Engineering and Information Technology

  Current (mid-funding): RMIT, School of Computing Technology

• 
  Dr Holly Chinnery, Senior Lecturer
  School of Optometry and Vision Services
  Faculty of Medicine, Dentistry and Health Sciences
• 
  Dr Vlada Rozova, Research Fellow in Applied AI in Medicine (Honorary)
  Formerly:

  School of Computing and Information Systems
  Faculty of Engineering and Information Technology

  Current (mid-funding): RMIT, School of Computing Technology