The Computational Discussion Continues…

In the first of this two-part feature (1), we discussed the impact and evolution of computational pathology. Now we ask: what can computational pathology do for you in the clinic? Many pathologists aren’t sure – and some fear that computers may impact (or even take away) their jobs. But those working to develop artificial intelligence technologies for pathology see it as an assistive tool, not a replacement.

Our computational consultants

Manuel Salto-Tellez is Chair of Molecular Pathology at Queen’s University Belfast, Clinical Consultant Pathologist at the Belfast Health and Social Care Trust, and Deputy Director of the Centre for Cancer Research and Cell Biology. A histopathologist and molecular diagnostician, he has studied and practiced internationally for the past 25 years. His current focus is on the integration of genotype and phenotype data for the improvement of personalized medicine and patient care.

Peter Hamilton is Business Lead for Philips Digital Pathology in Belfast, United Kingdom. Previously, he was Professor of Pathology Imaging and Informatics in the Centre for Cancer Research and Cell Biology at Queen’s University Belfast. His research interests include digital pathology, computer vision, tissue bioimaging, and the high-throughput analysis of novel tissue and cell biomarkers.

How can artificial intelligence and machine learning contribute to the laboratory?

Manuel Salto-Tellez: From the point of view of discovery and application in pathology, concepts such as tumor microenvironment, cancer ecology and morphomolecular heterogeneity are beginning to reveal to us just how much information in tissues and cells remains untapped because it cannot be seen by the human eye. We need to understand the diagnostic and clinical relevance of this information. New information frameworks, such as the ones provided by AI and machine learning, will be critical to the further exploration of these concepts, and to helping us understand how we can diagnose and treat diseases faster and better.

Peter Hamilton: Artificial intelligence (AI) and machine learning will provide a way to automate the analysis of the complex tissue and cellular images that pathologists look at every day. Philips has been developing advanced algorithms that can be embedded within our digital workflow to analyze the pixels in digital pathology images and use the data to identify a range of abnormalities. From a provider’s perspective, this has four major benefits for a digital pathology laboratory:

1. It can provide the pathologist with objective, reproducible data on the sample and thereby reduce diagnostic or prognostic variation, ensuring that clinical decisions are accurate and patients get the best care.
2. By implementing automated image analysis, some tasks can be accelerated, streamlining pathology workflows, removing inefficiencies and reducing costs.
3. Some diagnostic tasks may only be possible through AI; subtle changes in tissues, invisible to the naked eye, can only be detected using computer algorithms.
4. AI and machine learning can help integrate ever-increasing amounts of patient data – including pathology data – to establish individualized patient signatures and help drive precision medicine, targeted therapies and improved clinical outcomes.

What are the major pros and cons of AI in the clinic right now?

MS-T: The pros and the cons are both related to the volume of information. The amount of new data we can generate on a patient from multiple sources (genomics, pathology imaging, radiology imaging, clinical records, and even health information on social media) is phenomenal. Equally intimidating are the challenges involved in identifying the right algorithms to translate such a wealth of information into practical considerations. This translation arguably represents the most important hurdle to applicability. Crucially, AI and computational learning will help us analyze this massive amount of information to make it meaningful. With improved understanding, we can respond by developing solutions – and, hopefully, cures.

PH: AI-based algorithms need sufficient robustness to be able to handle the variation in sample preparation that exists between different laboratories. And that requires the use of extensive datasets to train and validate algorithms for clinical pathology. My colleagues and I are working with multiple partners, including pathology labs that have already moved to 100 percent digital workflows, to build some of the largest libraries of digital images across multiple disease types. This growing digital library of annotated images will help build robust AI for all of our future applications.

As Manuel says, the ability to integrate multiple sources of data is indeed a critical driver for AI in the lab. Pathologists are skilled integrators of complex information. AI can provide important tools that enable pathologists to do this more effectively across multiple data sources and multiple patients. The data management question definitely extends beyond pathology into other specialties where the level of available data – and the complexity of that data – is increasing year on year.

In one research study on the identification of lymph node metastases in breast cancer (2), the pathologist slightly outperformed the AI algorithm in correctly identifying metastases. However, when both the pathologist and AI analyses were combined, the overall accuracy rate significantly increased over what the pathologist was capable of alone. This example shows the real value of combining the skills of the pathologist with AI to reach a more precise diagnostic assessment of a case. I think it’s clear that hybrid machine/human intelligence has the power to drive major improvements in pathology in many ways.

Underpinning all of this is patient safety. Clinical and regulatory clearance for AI tools is of the utmost importance and is a very costly part of the process – probably the most expensive step. We have to perform the most robust clinical studies on our pathology AI tools and obtain full regulatory clearance before releasing them onto our digital pathology platform. That way, we ensure that pathologists can apply new tools with real confidence.

Is AI going to replace pathologists?

PH: To be clear, AI is not going to replace pathologists; rather, it’s going to enrich workflows and help pathologists better cope with the rapid rise in sample numbers and their complexity. We are currently working with large teams of pathologists to build exciting new tools – and many pathologists cannot wait to get their hands on them!

MS-T: Peter is absolutely right. Like many other past and present developments in pathology, AI is not here to substitute for pathologists, but to improve the quality of our diagnostic output. It is conceivable that the pathologist of the future will be involved in not only diagnostic interpretation, but also in the application of algorithms to images – and that the final diagnostic opinion may be a combination of both.

Can you describe an “ideal” situation for the application of AI and machine learning?

MS-T: In my opinion, digital pathology and AI can potentially be of help in five main areas of the “pathology diagnostic pathway.”

1. Finding a new way of managing images in routine diagnostics and in routine archival collections;
   
2. Integrating images with existing patient and pathological information more readably (LIMS, image reference resources, can texting, access to diagnostic bibliographical resources, and so on);
   
3. H&E interpretation and resolving areas of diagnostic dilemma (for instance, in situ versus invasive cancer);
   
4. Biomarker scoring, immunohistochemistry and in situ hybridization; and
   
5. Annotating samples ahead of nucleic acid extractions and molecular testing.

Thus, the ideal situation is “global” and, in many ways, limitless!

PH: There are many areas where AI can be used to support diagnostic reporting and reduce inefficiencies in pathology workflow. For example, it has been shown that deep learning AI can be used to rapidly detect lymph node metastasis in breast cancer patients with high levels of precision. Grading certain cancers using conventional diagnostic schemes and the human eye is also known to be associated with subjectivity and error. Computer-based AI and image analytics will certainly help ensure higher levels of consistency and precision across pathologists and better management for patients. In laboratory workflows, I can also envisage ways in which AI can intelligently manage workloads, dispatching cases to distributed pathology teams in ways that drive efficiencies, reduce costs and accelerate turnaround times. Such advances will be critical in overburdened pathology laboratories across the world.

What would you like pathologists and lab medicine professionals to know about AI in the laboratory?

PH: I would say two things. First, AI – as with any new technology in pathology – will really help enrich your ability to practice pathology and improve your ability to serve your patients. Second, AI is not here to remove pathologists from the decision-making process. It can and will help accelerate certain steps in pathology, but pathologists will continue to be responsible for the final decision. Pathology will change with the introduction of AI in the same way that it changed with the introduction of IHC, ISH, molecular pathology and now next generation sequencing – but, as with these advances, it will provide richer data on the interpretation, understanding and integration of morphology. I see this as an exciting step forward in the discipline of pathology – one that puts pathologists at the very center of clinical care and the delivery of precision therapeutics.

What kinds of equipment, software and training are required to implement clinical AI?

PH: To begin with, AI requires the implementation of digital pathology in routine diagnostic practice – without this, AI applications in pathology are not possible. As we see the adoption of digital pathology for primary diagnostics grow across the globe, AI is following close on its heels. Indeed, in many laboratories, AI and computational pathology are becoming key drivers for the adoption of digital pathology due to the efficiencies they can bring.

In addition, for AI to be truly effective, it needs to be seamlessly embedded within the pathology workflow. At Philips, we are building the technologies that will allow image AI to be an integral part of our existing digital pathology solution and diagnostic workflow experience. The goal? To make AI simple and accessible to pathologists, with results available at the time of review. We want to provide powerful decision support tools for pathologists to help them make the toughest clinical and therapeutic calls.

As with any new technology, pathologists will have to be trained in digital and computational pathology. It’s important to make that digital transition as straightforward as possible, providing technologies that enable digital pathology workflows rather than making them more complex. These include high-throughput scanners that have “walk away” technology and don’t require multiple technicians to run the instrumentation, image management and storage that is easily scalable, and user interfaces that have been designed by pathologists to enrich digital workflows and accelerate AI-based decision-making. Approaches like these reduce the level of training and support that pathologists will need to adopt digital pathology. The same applies to AI; the tools need to be easy to apply and provide results that can be easily interpreted and embedded into the diagnostic report. Having an existing, highly streamlined digital pathology workflow allows Philips to incorporate AI in a way that saves pathologists effort, something I hope will become standard practice as digital pathology becomes the norm. Pathologists won’t need to be data scientists or computer engineers to use AI; it will become a normal part of their diagnostic decision-making.

What advice do you have for pathologists moving to a more computational environment?

MS-T: Take the example of molecular diagnostics. Many tissue pathologists arrived late to this important area of diagnostics, and have thereby put it in the hands of others. Equally, not embracing AI from the beginning and allowing others to enter the field on our behalf may be devastating for the future of our discipline. If we successfully adopt AI ourselves, though, we will have engaged in a new diagnostic revolution – not dissimilar to the one Virchow led more than a century ago.

Every time new “glass” pathologists are involved in studies in which they become “digital” pathologists, they typically go through five phases: reluctant, skeptical, interested, involved, and enthusiastic. The “knowledge integration” that is facilitated by working in silico is an unequivocal advantage to almost every pathologist who gives digital pathology a fair try.

PH: I would advise pathologists and laboratory medicine professionals to work closely with industry to embrace change, and to ensure the reliable and effective implementation of new technologies in a way that truly enables best practices, processes, and workflows for your laboratory. It is equally incumbent on industrial vendors to truly partner with pathologists and laboratories to ensure successful digital transformation and the effective introduction of new AI technologies.

The only other advice I would give pathologists during this period of transition is to continue to be excited by pathology as a discipline and the future advances that are within our reach! Your role in defining that future is critical to pathology and to the patients you serve every day.

What’s happening right now with AI implementation in pathology?

PH: Philips is committed to developing AI applications for digital pathology users. We have also responded to an opportunity to work with the UK’s National Health Service (NHS), government, research organizations, other industries and the small and medium enterprise community to explore how we can apply AI to pathology to advance patient care in a fully digitized laboratory. In particular, we are looking at what we can offer to support the incubation and creation of world-class digital pathology hubs within the NHS, and to drive leadership in the transformation of diagnostic pathology. At the moment, we are working with the Office of Life Sciences and the Department for Business, Energy and Industrial Strategy to determine how best we can collaborate to drive improvements in NHS pathology and in AI research within the UK.

What’s next for computational pathology?

PH: Health knows no bounds. Philips conducted a recent global survey and found that 56 percent of people who had had a cancer-related experience felt that “connected care” technology made their experience more positive (3). We see digital pathology and AI as being cornerstones to connected care; they are keys to improving patient care and health outcomes. The digital transformation is happening – and will continue to happen – in medicine and in pathology. It’s nothing to be scared of, and it will be part of a broader digital revolution in health. It’s a genuinely exciting future – one that pathology and pathologists deserve.

MS-T: For the last 60 years, we have witnessed the “genomic revolution in medicine,” and we are now beginning to see tangible gains. I believe we are beginning to experience the “AI revolution in medicine,” a concept that encompasses genomic information and other sources of information – one that will provide a true “integromics” approach to healthcare.

To read Part 1 of this feature, please see our January issue.