"When I started in digital pathology, I felt the lack of a pathology-friendly open-source platform was a big problem." — Peter Bankhead
Since its launch, QuPath has become one of the world’s most widely used image analysis platforms for digital pathology research. Its impact on the international pathology community has been recognized by the Royal College of Pathologists, which recently presented its developers with a team award for innovation.
Yet, behind its success, QuPath's road to development has been beset with false starts, administrative battles, rejection, and frustration.
We spoke with QuPath creator Peter Bankhead, Reader at the Institute of Genetics and Cancer, University of Edinburgh, to hear what drove him to keep going against the odds.
What inspired you to develop image analysis software for digital pathology?
I hadn’t really heard of digital pathology before I joined the field as a postdoc in 2012. I’d developed retinal image analysis tools during my PhD, and I’d been an image analyst in a microscopy facility for three years, but I hadn’t appreciated how large and complex pathology images could be. The tools and techniques I knew weren’t designed for whole slide images.
I spent a couple of years trying to adapt image analysis tools to score IHC biomarkers in tissue microarrays. None of the plugins or scripts that I wrote were very successful, mostly because they were so hard to use. Still, the idea of developing a full software application to support whole slide images seemed so far beyond my abilities that I didn’t give it any serious thought.
How did QuPath come about?
Eventually, I decided to spend a few days writing a new whole slide image viewer, mostly to prove to myself it wouldn’t work. However, with the help of OpenSlide – an open-source library to read pathology image formats – I was surprised how quickly my new viewer started to look useful.
Then, once I had an image viewer, I started to add analysis features. Suddenly, I wasn’t constrained by adapting other tools. I could create the features I actually wanted and hadn’t seen elsewhere.
Eventually, this work became QuPath. It wasn’t planned, and it only happened because not writing QuPath hadn’t worked very well.
Were there any particular challenges or setbacks along the way?
On a personal level, I had to really fight to make QuPath open-source. In fact, that’s why I left my postdoc. After many months arguing the case for open-source, permission to release QuPath was granted only after I resigned, during my notice period. When my next job stopped me from continuing it, even in my spare time, I left that too. I spent some months unemployed, delivering a few talks and workshops, while trying to figure out what to do next.
Publishing QuPath was also difficult. My paper describing the software and its application to biomarker scoring was rejected by at least five different journals. It was considered unlikely to have much impact, and none of the editors sent it out for peer review.
This made for an exhausting and frustrating few years. The setbacks did, however, help me become less concerned about conventional measures of success, and more focused on my own ideas of what is worthwhile.
What do you know about QuPath users?
The main QuPath paper was eventually published in Scientific Reports. It has been cited well over 6,000 times, including more than 1,500 new citations last year alone. That underrepresents use, since many studies report using QuPath without citing the paper.
The software itself has been downloaded over a million times, but we don’t have data on exactly who is using it or how. Nevertheless, journal articles demonstrate that it is used all over the world to study a huge range of diseases.
It is used in both academia and industry, with several companies mentioning QuPath protocols or extensions on their websites. One major tech company even incorporated QuPath into product launch videos to demonstrate scientists using their computers.
How is the platform being used in the real world?
Many QuPath projects involve detecting and classifying cells within whole slide images of tissue. The classification might involve distinguishing between different cell types, using morphology, staining, or a combination of both.
QuPath’s application to Ki67 is especially well established: this involves identifying tumor and non-tumor cells, mostly based upon morphology, before calculating the percentage of tumor cells that are positive for Ki67, based upon staining. A few studies have independently compared QuPath with both alternative software and pathologist scoring for this purpose.
But it’s important to note that QuPath doesn’t have a dedicated Ki67 algorithm. Rather, it provides image processing and machine learning building blocks to create custom algorithms. There are endless ways to combine these building blocks, based upon whatever the user wants to do.
This makes QuPath flexible. It’s not limited to analyzing either cells or whole slide images. We currently have projects applying QuPath to both fluorescence confocal z-stacks and electron microscopy images.
Personally, I have used QuPath to help digitize a family photo album. I’ve heard it has also been used to look at fossils, and to quantify the fracture behavior of cheese.
How has the platform evolved over time?
We’re a small group – there has never been more than a few people working on QuPath at any time – but the user community is much bigger. There are tens of thousands of posts about QuPath online, spread across more than 5,500 separate topics on the Scientific Community Image Forum. It’s satisfying to see people helping one another, often giving far better answers than we could have. It’s also nice to see QuPath workshops being advertised all over the world, and students posting on LinkedIn about their QuPath projects.
Despite the large user community, we continue to write the core software ourselves. We try to make this code stronger over time, to enable our own research and that of others. If someone else wants to add a feature, we encourage them to create their own extension – and sometimes we create extensions ourselves.
For example, we collaborated with Joel Saltz’s group at Stony Brook University, especially Jakub Kaczmarzyk, to make their AI models easy to run in an interactive and user-friendly way.
We also created an extension for InstanSeg, which is a uniquely fast and accurate AI model to detect nuclei and cells in complex images. InstanSeg was created by Thibaut Goldsborough when he was a PhD student in my group. He made it so useful and portable that other developers have integrated InstanSeg into their software as well – which we encourage, because we aren’t trying to lock anyone into only using QuPath.
QuPath has also been used for teaching. I was incredibly impressed by the work of Aaron Yli-Hallila, a medical student at the University of Oulu, Finland. He created an entire open-source teaching platform called OpenMicroanatomy, which includes a QuPath component called QuPath Edu. OpenMicroanatomy and QuPath Edu have been used to teach medical students in Finland for several years, with further pilot studies in South Africa and Namibia.
How do you see it expanding in the future?
Currently, there are two of us working on QuPath’s code: myself and Alan O’Callaghan, a research software engineer and postdoc in my group. Alan knows a lot that I don’t, and we both know things that I didn’t back when QuPath started. Together, we’ve amassed a collection of ideas to make QuPath a much stronger platform in the era of AI, multiplexed, and multidimensional imaging. We’ve got a year or two before our funding runs out, so the race is on to implement these.
Why is it important to you that QuPath is available as an open-source platform?
My career wouldn’t have been possible without open-source software. I used ImageJ during my PhD, and then almost daily when I was an image analyst in Germany. I ended up writing an open handbook about it.
When I started in digital pathology, I felt the lack of a similar pathology-friendly open-source platform was a big problem. I regularly read papers describing the use of in-house software or proprietary tools. It wasn’t possible to verify the claims, reproduce the results, or apply the methods to our data. This seemed incredibly – even unethically – inefficient.
I could understand why companies didn’t make their software open and freely available to everyone. It’s hard to make a business case for that, and "free" software isn’t free to create.
I also understand the reticence toward sharing in academia. It’s scary and exposing. What if people find bugs? It’s also difficult and time-consuming to write and document software to a standard others can use. This extra effort is entirely counterproductive if you want to publish frequently and in leading journals, so it isn’t a very shrewd career move either.
Therefore, my initial surprise that digital pathology didn’t have an open platform dissipated when I considered that no one seemed to have any incentive to build one. If anything, it was actively disincentivized.
Still, it bothered me to think how many other PhD students and postdocs there must be in a similar position to me. How much time and research funding were we wasting to reinvent wheels that no one would ever use again?
That’s why I was determined that any software I wrote should be open-source. I wasn’t sure whether it would be any good or anyone would use it, but I had no interest in commercialization and couldn’t justify creating another in-house tool.
Would you like to see QuPath, or a derivative, being approved for clinical use in the future?
No. I think that many computational pathology groups and companies are aiming toward clinical use of their tools and software. If a problem already gets a lot of attention, then I’d rather spend my time on something else.
Trying to adapt QuPath for clinical use would inevitably make it less flexible and useful for research. It would also introduce legal and regulatory considerations that I suspect companies are better placed to handle.
That isn’t to say that I don’t want our work to help patients or pathologists. Rather, I think we do that far more effectively by maintaining our current priorities and values.
How do you view QuPath's impact on research?
Although it’s hard to quantify the impact of QuPath’s use across thousands of studies, I think it’s reasonable to assume it has helped labs worldwide make substantial cost and efficiency savings. QuPath has made some studies possible, and others unnecessary. I hope we’ve also had a positive influence on research culture in digital pathology, by helping reduce the barriers to openness and sharing.
Secondly, QuPath should improve the quality of algorithms in this field. If QuPath addresses a problem, then every later tool created to address the same problem should be better. If it isn’t, then what’s the point? QuPath is freely available to everyone, and we’ve shared all our code, so anyone can see exactly what we’ve done. It makes little sense for anyone to publish something new and worse.
It’s liberating to realize that, with open software, we aren’t competing for users. Instead, we’re trying to help raise the baseline to benefit everyone. Even if this doesn’t translate into recognition or revenue, it’s quietly satisfying – and that’s enough.
To what extent should we balance excitement over AI-augmented image analysis with caution?
AI presents us all with a lot to think about. I believe that the most interesting and important questions are not just about what the technology can do, but also about how it is used and who it will benefit. While I certainly hope that any use of AI in clinical practice will benefit pathologists and patients, I don’t think it’s inevitable. I’m glad to see more studies into topics such as automation bias. There is a huge difference between a proof-of-concept published in a journal and a genuinely useful software tool, so it’s important not to be too easily impressed – especially if the paper describes methods that aren’t made available in a way you can try.
So perhaps the exciting thing is that the role of digital pathology is still being worked out, which means it can still be influenced. If you’re a pathologist, I hope you won’t trust what computationally minded people like me claim our tools can do. Rather, I hope you’ll try them out where you can, engage critically, ask awkward questions, and help shape how the field continues to develop.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
