Counterfactual Visual Explanations in Ophthalmic Imaging

Lead Research Organisation: University College London
Department Name: Institute of Health Informatics

Abstract

While AI now betters the best humans in games like chess and Go, human-AI partnerships still
beat the best AIs in such games even with only a moderate human player. We believe this
paradigm will extend to AI-enabled healthcare and so have a broad aim to design AI systems
specifically to support human decisions.

This project exploits new ideas emerging in computer vision to design novel systems that
enable human-AI partnerships in the assessment of medical images. Specifically, it will
implement a system that provides counterfactual visual explanations (Goyal ICML 2019)
enabling human dialogue with AI systems that assess optical coherence tomography (OCT)
images of the retina via a process of image synthesis (creating an adapted version of the
image).

We explore a highly innovative and recent AI solution - counterfactual visual explanations - in
the context of healthcare for the first time. It offers great potential in pressing healthcare
challenges by enabling better human-AI partnerships to enhance overall decision making. The
project nicely matches the stated aims of the centre:

1) Extracting more information from patient data to accelerate diagnosis:
The huge collection of OCT data from patients at Moorfields is a fantastic resource with
already demonstrated potential to accelerate and improve accuracy of diagnosis; it
provides the perfect exemplar for maximising the potential of human-AI partnership.
2) Creating adaptive and flexible systems that improve the operation of healthcare
organisations:
The human-AI partnership paradigm that motivates this project addresses this aim
directly and has much wider potential beyond the demonstration in ophthalmology.
3) Delivering personalised and targeted treatments for patients:
Better diagnostic and referral decisions directly enable personalised treatment and care
design

The image synthesis system created as part of this project will also have wider potential in data
augmentation, system validation, and human training systems.

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/S021612/1 01/04/2019 30/09/2027
2245851 Studentship EP/S021612/1 23/09/2019 30/09/2023 Peter Timothy Woodward-Court