Development of Artificial Intelligence OCT Biomarkers for Accelerated Skin Disease Research and Diagnosis

Our vision is to bring the power of machine learning and computer vision (also known as 'Artificial Intelligence' or AI) to the application of Optical Coherence Tomography (OCT) imaging of skin, in order to dramatically improve the speed, accuracy and utility of these OCT imaging devices to dermatologists and clinical scientists.

At present, end-user clinicians and scientists use OCT imaging devices to capture sub-surface images of skin and then they manually analyse the images to extract data, which is then used to assess the effects of pharmaceutical treatments on skin diseases. OCT imaging is faster, less invasive and less costly than taking skin biopsies, but the image analysis step is still time-consuming, somewhat subjective, and requires observer training. This is a hindrance to the use of OCT imaging to accelerate drug development for skin diseases like skin cancer, atopic dermatitis and psoriasis, which are multi-billion-$ markets.

We believe that powerful machine learning algorithms will transform how OCT skin imaging is used by clinical scientists and clinical users to research and develop new drugs.

To achieve this vision, we propose seconding a leading expert from AI specialists Manchester Imaging Ltd (MIL) to the host organisation Michelson Diagnostics Ltd, UK SME manufacturer of the world-leading VivoSight Optical Coherence Tomography (OCT) skin imaging and measurement system, over 2 years, to develop and test novel machine learning algorithms for OCT.

Key barriers to the wider adoption of OCT for dermatology research is that the OCT images require trained experts to interpret them, and also that the image analysis is somewhat manual in nature. Dermatologists are often time-poor and may not have time to learn how to do this, and the manual nature of the analysis creates potential for unwanted bias.

Therefore the challenge is to reduce the barriers to adoption by:
Automatically identifying image-markers for common skin diseases
Automatically quantifying the image-markers

Examples of OCT image-markers requested by Michelson's user base are:
Thickened epidermis (Atopic Dermatitis)
Loss of definition of dermis-epidermis junction (skin cancer)
Detection of tumour 'nests' in the dermis and their invasion-depth/extent (skin cancer)
Increase in blood vessel density (all inflammatory diseases)
Alterations in blood vessel shape/tortuosity (melanoma)

The challenge can only be met by bringing together expertise in AI-algorithms (image processing/machine learning) and OCT imaging technology (laser physics, optics and instrumentation) with close links to the end-user clinical science user base, to form a highly focused and motivated multi-disciplinary team and who will develop and test candidate algorithms on real clinical data.

The research aims to use AI techniques to develop novel biomarkers for OCT skin scans that will aid end-user interpretation of the scans thus providing added value and utility for the scanning devices. Drawing on an existing library of over 500 scans, including detailed clinical reports, we will use image processing and machine learning techniques to train systems to identify image biomarkers. These will be presented to end-users in a form that will facilitate their interpretation of the scans and allow them to make better informed diagnosis, disease monitoring or treatment planning. The project will deliver working prototypes of AI systems that identify and display 4 novel image biomarkers for OCT scans of skin together with detailed performance evaluation data and release and maintenance plans suitable for their integration into a medical device.

Phase 1 of the project plan involves detailed analysis of the technical challenges; understanding the end-user requirements to ensure the proposed solution is relevant; appreciation of background IP to benefit from previously developed and published solutions. Phases 2 & 3 employ rapid prototyping techniques, common for AI system development, to provide early evidence of component efficacy and a regular stream of usable deliverables. Examples of machine learning methods which could be deployed are deep learning classification to identify parts of the scan that are indicative of particular disease or 3D active shape models to delineate structures of interest. The prototypes are useful for internal evaluation and demonstration to customers. Included in the plan, but not costed as part of this project, are work packages assigned to Michelson to provide datasets of images and annotations necessary for the training and evaluation of the AI system. Phase 4 focusses on delivering detailed evidence on performance and efficacy of the prototype systems and a plan for integration with existing Michelson medical device systems.