Paving the Computational Path

How did you get into digital pathology?
I was visiting a collaborator in the Yale University applied mathematics department in 2002 when he shared some multi-spectral images of colon histology sections for me to look at; It was fascinating to see nuclear and subcellular level details in those images. I did some initial image analysis on them using texture and morphological descriptors and found it amazing that one could differentiate between normal and cancerous features based on digital descriptors. I have been glued to pathology images since. After my return from the US, I initiated collaborations with other colleagues in the UK and Germany to work on pathology images myself before digital scanners were available – and before it became known as “digital and computational pathology.”

In 2008, we organized a computational histopathology workshop at the International Symposium for Biomedical Imaging; to the best of my knowledge, it was the first international meeting on computational pathology (1). A unique opportunity then came along in December 2010; a US slide scanning manufacturer looking for a demonstrator site and center of excellence in the UK chose Coventry and Warwickshire Hospital (the teaching hospital of the University of Warwick), at least partly due to our existing humble track record in digital pathology. That led to what still remains the world’s largest validation study on digital pathology for primary diagnosis, published in Histopathology 2016. The Coventry pathology lab became the first UK hospital pathology laboratory to do live reporting on digital slides and I was fortunate enough to be involved in that stellar effort, led by Ian Cree, from day one. That was a real turning point for us because it generated tons of imaging data being digitized on a daily basis. Today, most of the histopathology reporting in the lab is done digitally.

Over the last few years, we have focused on gaining insights from the billions of raw pixels in each of the digitized tissue images for large cohorts of cancer subtypes to discover new markers for predicting survival, clinical outcome and response to therapy. And we are now developing machine learning algorithms that can reveal histological signatures that may not be recognizable by the naked eye, with a view to guiding diagnosis and treatment.

What is PathLAKE – and what do you hope to achieve?
PathLAKE is a national center of excellence for artificial intelligence (AI) in pathology that will house a large-scale data lake – a centralized and accessible repository of clinical pathology imaging data for AI researchers. It involves sourcing high quality annotations from pathologists, which will be made available with the pathology images together with linked clinical metadata. Such resources will be unquestionably key for training high-quality AI algorithms for further downstream analysis. A unique aspect of PathLAKE is that it not only allows you to access data for research and AI innovation purposes, it features an in-lake central analytics engine to allow you to conduct experiments within the data lake; for example, if you don’t have sufficient computational resources or storage.

How will accessible data affect the adoption of digital and computational pathology?
I think making large data repositories available to researchers and industrial players will be critical in further driving AI innovation. It will help accelerate the development of algorithms that will solve some of our most fundamental problems and optimize their solutions. It will also help us to make new discoveries – allowing us to stratify patients into previously unknown subgroups.

Right now, it’s a little early to see the direct impact of computational pathology on patient healthcare, as I don’t believe many algorithms have been adopted in clinical practice. But it is coming. Several hospitals in the UK are now in the process of being digitized. The recent £13 million awarded to PathLAKE Plus will further digitize several more hospitals in the UK and road-test some of the AI solutions in a clinical setting.

The COVID-19 pandemic has been a bit of a diversion from our main goals, but we are coming up with a COVID-19 workstream that will allow us to input various types of data so researchers can start exploring the datasets. Once we’re able to use machine learning to understand COVID-19 from various angles, we can develop new algorithms to further understand its beastly nature.

What does the future hold for computational pathology?
I believe the key game-changer will be data repositories that simplify the development of algorithms for machine learning researchers.You won’t need to put together your own datasets, which requires a huge effort in itself, before you can develop cutting-edge algorithms that can push the state of the art and make translational impact. That’s the philosophy we are following with PathLAKE. There’s so much that can be achieved when you make data and computing resources available to smart, young people who are quickly bringing themselves up to speed with the latest deep-learning technologies.

What advice would you give to trainees who have an interest in computational pathology?
Try to be brave through the initial daunting experience and don’t hesitate to jump into what may seem like an unknown territory. It’s only by crossing the boundaries of disciplines that you are going to make a true difference and impact. Some of the best ideas come from people working in other domains that bring new perspectives into the field. These people are future leaders – not the ones who are comfortably sitting in their own space and sticking to what they are happy with, but those who are willing to step outside of their comfort zones.

And a take-home message for pathologists?
Change is coming your way, but it’s change for the better. Embrace it and be adaptable rather than resisting it, because digital and computational pathology is the future. We are all going to have to live with that or else risk becoming dinosaurs.