20 Years of Digital Pathology: A Conversation with a Pioneer
Twenty years before AI became the hottest topic in healthcare, some researchers were already asking a fundamental question: what if a computer could understand a pathology slide?
Today, digital pathology and artificial intelligence are accepted methods used in the diagnosis of cancer for patients around the world. Whole slides are digitised in minutes, and algorithms are used to support pathologists in making faster, more consistent decisions. However, in the early 2000s, this technology was in its infancy.
Commercial whole-slide scanners were only just beginning to emerge, and the idea of routinely analysing digital pathology images was still, for the most part, experimental. However, researchers like Cirdan’s own James Diamond were already building the foundations for what would become a significant technological advancement in healthcare.
In 2004, Jim authored and co-authored various pieces of research, in an attempt to make the leap from the traditional glass slides method, to digitised images in pathology. In one piece of co-authored research, he described a ‘machine-vision’ system, capable of analysing digitised prostate tissue, using morphological and texture analysis, and generating digital maps of tissue composition. The paper represents one of the major early milestones in computational pathology: at a time when digital pathology was still emerging as a concept, it demonstrated how digitised histological images could be analysed computationally to distinguish different tissue types, anticipating many of the concepts that underpin today’s AI assisted pathology systems.
The current excitement surrounding AI can sometimes make innovation appear sudden, when in reality, today’s breakthroughs rest on decades of incremental research by pioneers like Jim, and a constant evolution in engineering and clinical collaboration.
Today, Jim forms an integral part of Cirdan’s AI team, as a Senior AI Validation Engineer. We caught up with Jim to find out more about his experience of the early days of digital pathology and AI, and how far the discipline has come since then:
What did pathology look like in the ‘90s and early 2000s, and what convinced you computers could play a role?
I bought my first PC in 1993 for £1100, to write my PhD thesis. It had 1Mb of RAM (Yes, you did read that correctly). I upgraded the hard disk to 80Mb; Windows 3.1 came on 6 floppies. At this stage, the specs of the average NHS computer were significantly less than this.
At that time, our work in the Quantitative Pathology Laboratory in the Royal Victoria Hospital, Belfast, mostly followed that of pathologists. They carried out their daily work, and we followed along to carry out the quantitative work afterwards. We had cameras to attach to our microscopes, of which the highest resolution was 262 Kilopixels. Digital cameras were not available yet! One arrived in the year 2000, bringing us up to 1.2 Megapixels.
We were aware that imaging was going to be relevant to the future of Quantitative Pathology. I spent the first six years of my post-doctoral career attempting to measure the things that pathologists estimated by eye. Papers were published on simple hypotheses “are tumour cells larger than normal cells?”, “variation in boundary profile”. All of this work was conducted on grayscale images of 512 x 512 pixels. Some of the researchers wanted to push further, and that meant working on images that were much larger than a single field.
Before commercial whole-slide scanners existed, how did you actually create digital images large enough to analyse?
In radical prostatectomy specimens (about the diameter of a golf ball), I, in more recent years, have scanned and produced JPEG images which were roughly 250k x 250k pixels. Images of this size were simply beyond comprehension to researchers back in the late 1990s. To compare, in the ’90s I had access to one single channel 512 x 512 pixel camera. This worked well for the work we were doing in nuclear morphometric analysis. However, we wanted to do some work in the lab that would push research into a more ‘diagnostic’ arena.
I decided that I could begin by writing a paper on tumour identification, specifically with morphologies that were not possible with a simple microscope-mounted camera, simply because they were much larger. This paper was published in Human Pathology, 2004. Based on preliminary research, I had estimated that I needed images that were around 20MP. This was really testing the limits at the time, considering the average image size was around 250KP! I asked our technician if this was possible, who made the mistake of saying ‘yes’. I got to work and proceeded to acquire an array of images, which were manually stitched together to form large scenes. These images, in 1999, were enormous. However, my research had one fatal flaw I had forgotten to assess: ‘Do I have any software that can open a 20MP image?’ For context, my phone’s camera is currently 33MP.
Just after the Christmas holidays of 1999, I started using a brand-new, shiny application brought about by two PhD students (Google), to search for any work on creating large pathology images. The term Virtual Slide had not been coined yet. Here, I found a start-up operating from a garage in California (Aperio). I contacted the owner of this new organisation, and sent him a slide (possibly the first NHS slide ever scanned, 2000). He scanned it and posted to me an image 50,000 x 30,000 pixels (no downloading here!). There was absolutely no software available to open or analyse this image. So, I wrote a piece of code to section it into 5k x 5k squares in C++, knowing that I could analyse them. This may well have been the first scanned image to be analysed in a publication!
When you published your research in 2004, did you imagine that AI would one day become part of routine pathology practice?
I suppose I did know, but I knew in my mind that it was a long way off. The term ‘Artificial Intelligence’, or ‘AI’, is not what we understand now. Then, I knew nothing about AI models, loss functions or hyper-parameter tuning. AI was beginning a transformation from rule-based “expert systems” to machine learning. The paper in question did not use AI as we know it now, but rather system of rules and measurements from images that could help quantify the morphometries of tissue that could identify tumour. AI today is orders of magnitude more than what I could have imagined 25 years ago. There was a strand of research looking at the concept of Neural Networks (NN) but these were frowned upon by the pathology community, because evidence could not be presented as to why the “computer says yes/no!”
Looking back, which part of the research are you most proud of?
Call me old fashioned, but I have always considered that the ultimate output of any of my research is its impact in the pathology community, and ultimately patients, rather than being for myself. My proudest moment was when the Thames Pathology Consortia bought my software to train all their junior members. This was a truly pivotal point for me, to know that my years of research were making a tangible difference for the better, in industry.
What do you think the future holds for Digital Pathology and AI?
AI has a huge future and a lot of trust to earn. Everything currently is about workflow and getting pathology departments and pathologists to be efficient, in a world of increasing demand. For example, a pathologist can have a fair idea of a particular diagnosis in 5 minutes, but spend the next 25 minutes reporting. Digital Pathology and AI innovations are consistently working towards making this process faster and more streamlined, while maintaining accuracy. This will be how AI moves into general pathology use. It’s great to see the next generation in Cirdan, in both Research and Engineering, taking all the ideas forward into a brave new AI world.
More than 20 years after this research was published, digital pathology has become an integral part of modern healthcare, with AI helping to shape the future of cancer diagnosis around the world.
At Cirdan, we believe that understanding the journey of emerging technologies is just as important as embracing the opportunities they create. By learning from the pioneers who helped shape today’s innovations, we can better prepare for the breakthroughs of the future and continue supporting our clients in a rapidly evolving technological landscape.
Thank you to Jim Diamond for taking the time to share his experiences and reflections with us. His insights offer a fascinating perspective on a pivotal moment in the evolution of digital pathology and highlight the importance of looking beyond today’s challenges to imagine what might be possible tomorrow.