Enhanced phenotyping of melanoma whole slide images with artificial intelligence

Research Context
Cutaneous melanoma is curable if resected early but higher stage tumours commonly behave aggressively. Melanoma continues to increase in incidence. Currently, all diagnoses are made by histopathologists. This assessment is highly subjective and in difficult areas like melanoma diagnosis, there is wide variation in opinions between histopathologists, resulting in a large proportion requiring specialist review and 1 in 5 diagnoses being revised. Moreover, unfortunately there is often also disagreement between specialist dermatopathologists.

Recently, new technologies have enabled the 'digitisation' of stained tissue slides, allowing histopathologists and researchers to view the tumour tissue at a cellular level on a computer screen. This has allowed the use of artificial intelligence (AI) algorithms to provide objective assessments of the tissue. We hypothesise that the application of these technologies may be able to provide a second histopathological "opinion" (both more objective and reproducible) to improve the accuracy of diagnosis and prognosis estimation in melanoma, with the potential to identify new prognostic biomarkers.

Pilot Work
Through collaboration with experts in melanoma tumour biology, digital pathology and computer vision, as well as dermatologists, mathematicians and dermatopathologists I have been able to develop and configure AI algorithms to a proof of concept sample of melanoma digital slides. We have a large pre-existing dataset of melanomas of a range of subtypes from the Leeds Melanoma Cohort (n=2184) which has comprehensive pre-existing clinical and somatic genetic data, which is ideal for training AI algorithms.

Aims and Objectives
The aims of the research are to: (1) develop AI algorithms for the extraction of morphological features at a cellular level to assist the histopathologist in their assessment by providing an automated, reproducible and objective second "opinion" for improved diagnostic accuracy and prognosis estimation and (2) explore the use of these morphological features to generate new prognostic biomarkers.

To achieve these aims, we have identified the following objectives:
1. To derive an accurate, objective and reproducible method of analysing the melanoma phenotype through detection and sub-classification of cells within melanoma whole slide images of a range of subtypes. The data derived will be validated against manual annotation.
2. To visualise and analyse this quantitative morphometric data in 3D for improved understanding of tumour shape in relation to surrounding structures and to test physiological models of the key drivers of melanoma metastasis by comparing the data to computer-generated models of tumour growth.
3. To analyse the data with respect to specified tumour driver mutations (including BRAF/NRAS, p16, kit, NF-1) and survival (time to death).

Potential Applications and Benefits
We postulate that this novel research will lead to the development of the technology for use in analysing the appearance of melanocytic lesions in clinical practice to provide more accurate diagnoses/prognosis estimation and therefore prevent patients from being over or under treated. The general principles established will further inform the development of similar tools for other pathologies. Second, this technology permits the generation of objective morphological data that cannot be captured by other means, which may provide new insights into how tumours develop, organise and spread. It will also help us understand how tumour cells interact with the body's normal cells (including the surrounding skin as well as the patient's immune system) and how the underlying somatic genetic drivers impact the on the growth patterns of the tumour. Ultimately, this project aims to deliver increased understanding of the melanoma pathology to help provide more information of increased quality to clinicians for the benefit of patients.

Background
The pathology of melanocytic lesions is one of the most difficult diagnostic areas, resulting in high levels of diagnostic discordance. This project involves the use of convolutional neural networks and random-forest algorithms to afford a reproducible, objective and quantitative assessment of the melanoma phenotype.

Aim
In this proposed research, we aim to: (1) develop and configure algorithms for the extraction of morphological features at a cellular level to assist the histopathologist in their assessment by providing an automated, reproducible and objective second opinion and (2) explore the use of these morphological features to generate new prognostic biomarkers and better understand melanoma growth patterns.

Objectives and Methodology
We will employ machine learning methods to generate quantitative cellular data by using a large dataset of whole slide images including atypical melanocytic tumours and melanomas. 3D reconstruction will be performed on a subset of cases to enable improved visualisation and analysis of morphological data as well as comparison to computer-generated models of tumour growth. Data will be correlated with survival and somatic mutation status.

Opportunities
This novel research will enable the derivation of previously uncaptured phenotypic metrics, which may improve the existing subjective histopathological assessment byproviding an automated second "opinion". It is hoped that this approach will increase diagnostic accuracy in this notoriously difficult area to ensure patients are treated appropriately. These new data may also enable the discovery of new prognostic biomarkers through improved tumour subtyping and through comparing this morphological data to the genomic parameters, we may be able to detect subtle morphological distinctions that indicate the underlying mutation, resulting in improved understanding of the interaction between genotype and phenotype.

Melanoma is a relatively common malignancy with a rising incidence. Advanced disease is amongst the most aggressive and therapy-resistant cancers, unaided by it also being notoriously one of the most difficult diagnostic areas in histopathology, resulting in particularly high levels of diagnostic discordance between histopathologists. The use of artificial intelligence algorithms, has the potential to objectify and quantitate phenotypic features making morphological assessment more accurate and reproducible, whilst providing potential new prognostic biomarkers. Our vision, is to use new technologies to reveal previously unknown morphological distinctions to advance our current subjective phenotypic assessment for increased reliability and improved prognostication in melanoma. We anticipate that this work will be highly beneficial to the following stakeholders:

Patients with cutaneous melanoma: The findings of this fellowship, combined with the anticipated subsequent surge in research in this field, will likely be hugely beneficial to malignant melanoma patients in many ways. Firstly, through improved morphological analysis for increased diagnostic accuracy. This is essential to ensure patients receive the correct treatment. Secondly, the quantitative morphological assessment may provide new prognostic biomarkers which may help to better sub-stratify patients for treatment.

Myself: This fellowship is essential to develop my career as an academic dermatopathologist. More specifically, it will allow me to have a formal research education, whilst increasing my understanding of malignant melanoma histology, genomics, computer vision and data analytics.

The University of Leeds: This work will cement the University of Leeds as a leading centre of digital dermatopathology through the combined expertise of the melanoma and digital pathology research groups.

The University of Glasgow: This research will further the University of Glasgow's reputation and the foremost UK centre of cellular behaviour and the environment in cancer metastasis.

Wider academic community: We anticipate that the dissemination of this work will promote others to start using similar techniques, which will be hugely beneficial for the wider academic community and research area for melanoma and for other cancers. Moreover, my position as a ACF in digital pathology is unique in the UK. Through the presentation of my work at pathology conferences in particular, I hope to promote interest in digital pathology and encourage pathology trainees to get involved; pathologist's involvement in development and application of digital pathology is essential to ensure that the opportunities this technology provides are appropriately applied. Moreover, we are in the process of seeking ethical approval for sharing of the anonymised Leeds Melanoma Cohort with other academics; this will be a highly valuable resource for other academics.

NHS: The use of AI to develop diagnostic algorithms for melanoma would lead to its application to other areas of diagnostic difficulty and potentially to centralised high-throughput screening in histopathology. We postulate that this would be very beneficial to the NHS, particularly since patient populations are expanding rapidly, whilst pathologist numbers are dwindling.

Industry: The health technology industry who provide the artificial intelligence software will benefit from this research, as the software will be tested and further developed through the course of this project. Other companies will also be aided by highlighting the potentials for this technology.