In silico pathology - from bench to bedside

The DoMore! lighthouse project has used AI/deep learning to develop methods for automated analyses of pathology samples for improved cancer diagnostics. The methods have been scientifically validated, but further testing in clinical settings is a vital part of our commercialisation strategy. In particular, the systems need to be invariant to different scanner platforms and to variations in staining, and the health economic aspects of our methods need evaluation. The Norwegian Research Council has provided an extra grant of 5 million NOK for these purposes.

Implementation of the commercialisation project is required to take the promising methods beyond the research laboratory and to the patients. The goals for this project are to establish:

1. a prospective CRC trial to validate clinical utility; (Work Package [WP] 1)
2. software productisation including integration with existing clinical systems; market analyses, including cost-benefit considerations, and establishment of communication channels to relevant target groups; [WP2]
3. the development of a standardisation method to provide scanner- and laboratory invariant measurements; [WP3 and WP4]
4. decision support systems to integrate new biomarkers with established predictors [WP5]

WP1 - Prospective trial

Due to the Covid-19 pandemic, it has not been possible to initiate the planned prospective trial. We have instead planned for a Clinical Feasibility Test involving five hospitals in the UK, each being offered ten free Histotyping CRC tests through the University of Oxford. The study will record to what extent the test alters the treatment of the patients tested. We are currently awaiting formal approval from Oxford.

WP2 - Productisation, design and dissemination

Productisation

A prototype of an Histotyping CRC application is built for the Windows platform and shown to give the same classification results as the UNIX based system. On a “gamer-PC”, running cases in batch mode require about one minute per case.

Health economic study

Economic evaluation serves a strong role in healthcare and in many countries is compulsory for considerations of reimbursement. Serving as an analytic and advisory tool, economic evaluation aims to provide an assurance of value for resources used and enables transparency in decision making. Economic modelling provides a structure in which to analyse costs and effects of a health intervention to guide decision making.

With Histotyping, if we can accurately, and in an automated fashion, categorise patients by their likely outcomes and assign them to the appropriate treatment pathway, there is undoubtedly much to be gained. To make this justification, we need to demonstrate the benefit both to patients through greater access to a tool that can reduce their treatment burden and increase their quality of life, and to the healthcare system through a reduced cost, and a broadened provision of care.

In order to do so, we will apply a decision analytic model to simulate different scenarios and changing input parameters based on prognosis, treatment and effects for colorectal cancer patients. To ensure a robust modelling outcome, we will follow internationally recognised good modelling practices set forth by ISPOR. Additionally, we have enlisted an expert advisory panel for model input scrutiny and quality assurance.

WP3 - Data collection

Ten sets of 50 unstained slides were distributed to 10 laboratories, where they were HE stained and scanned, both XR and AP. The ten laboratories were

• Pathology OUS
• SiV
• Cheltenham
• UCL
• HCA Laboratories
• Liverpool/Blackpool
• Utrecht
• Amsterdam
• Unilab
• SCOT

A further ten sets of 50 HE-stained slides were sent to scanner manufactures:

• Zeiss
• Axio
• Philips
• 3D Histech
• Leica
• Aperio
• Hamamatsu
• Ventana
• Omnyx (Inspirata)
• Olympus

WP4 - Development of image standardisation method

The initial project plan involved the implementation of established methods for image standardisation, such as the method for standardisation of imaged HE-slides by Macenko et al., 2009. This approach aims to remove laboratory-procedure-specific staining effects and scanner-specific-imaging effects and make the imaged slide look the same regardless of where it is stained and on which scanner it is imaged. A calibration slide with reference colour intensities is studied and may prove useful in the work. During 2020, work on the Perspective article for Nature Reviews Cancer revealed the potential of using image augmentation and distortion to make the method robust enough to such variations. Instead of modifying the image, this approach makes the method invariant to these effects. This is particularly relevant when the stain- and scanner variation is not present in the development dataset, which is often the case.