Game Winning Hits: Singles or Home Runs?
Posted 29th June 2018 by Jane Williams
To use a baseball analogy – there have been more games won with singles than with home runs. Home runs, when they happen, are spectacular, but they are few and far between. The journeyman single, while not the stuff of legend, gets the job done, and happens frequently. When it comes to the practical application of artificial intelligence in digital pathology, the analogy holds. The home run is machine diagnosis – replacement of the pathologist – the holy grail. The single represents baby steps which improve pathology practice, right up until the technology reaches its inevitable maturity.
A lesson can be taken from the field of automated cervical cytology in the 1980s through the early 2000s. Cervical cytology practice is a tedious search for needles in haystacks – rare abnormal cells in a sea of normal cells spread across the surface of a slide. Automated methods of analysis had been sought with much money spent beginning in the 1950s. Initial ideas put forth were for a “home run” – automated diagnosis with little or no human decision-making – but after 30 years of trial, it became clear that the task was too difficult, and the automated diagnosis was just not going to happen.
Improvements in technology
Efforts were retooled to look for ways to use all the technology that had been developed with a more practical approach – the “single”. A device that gave probabilities of slides abnormality was the outcome. These instruments were then used to identify highest risk slides, first for directed (and more fruitful) quality control, then for identification of lowest risk slides which did not need to be reviewed by humans at all, and subsequently leading to localisation of potentially abnormal cells. These instruments were widely integrated into practice and remain at work in these roles today.
In hindsight, these sequential improvements in technology and their application in practice makes perfect sense. The continuously building of more complicated and improving processes allowed the Food and Drug Administration, in its regulatory capacity, and users, in their consumption of the technology, the opportunity to understand and therefore evaluate the effectiveness and value of the products being developed and introduced.
Most recently, automated cytology continues its evolutionary process using score generation combined with HPV typing data. This technology promises to move cervical cancer screening to the next level, generating results that show that have the potential to evaluates a patient’s risk of cervical disease, and hence guide their management. Together, the combination can do so as well as a human can evaluate a cytology specimen.
In 2018, artificial intelligence applications can learn a lesson from this brief history detailing the path which automated cytology took on the road to regulatory approval and acceptance by practitioners. Digital renditions of histology slides allow the application of machine-learning technology in a variety of forms. Devices/software can review and learn from annotations made by pathologists or they can analyze specimens with known outcomes to make their own conclusions about what each new presented image represents. Is it like the annotations? Is it like the outcomes? And with what probability of success? There will always be false-positive and false-negative outcomes, and hence with probabilistic assessments, there will still be a place for human intervention and final interpretation – right up until that eventuality of fully automated machine diagnosis is achieved.
Acceptance in the marketplace
In addition, improvement in the “machinery,” such as human/device interfaces being made more efficient and user-friendly, will constitute parallel baby steps which will aid acceptance in the marketplace. Such applications must follow a path – one that the regulatory agencies can use to build experience and one with which potential users can become familiar and confident. So the “home run” of an immediate path to machine diagnosis will most likely give way to the “singles” of productivity and triage using “pieces” of the technology as each is developed along the way.
We are already seeing such applications coming to light – automated and potentially more accurate counting procedures, such as:
- Immunohistochemical & in situ hybridization scoring
- Registration of slides for simultaneous multiplexing digital images of multiple special stains
- Machine localisation of glomeruli on medical renal biopsies
- Prescreening of a variety of biopsy types including prostate and lymph nodes, for the presence of high probability hotspots potentially representing cancerous involvement
- Improvements in interfaces that allow for more efficient and less tiring human/device interactions.
What is the solution?
Each of these processes does not provide a definitive automated solution – yet. Each provides a definitive productivity enhancement for the practitioner. A prescreened cervical cytology slide with areas of potential abnormality already identified can be accurately evaluated much faster. It would be expected that a prescreened prostate biopsy or lymph node could also be accurately evaluated (by a human) more rapidly, as could gastrointestinal biopsies, liver biopsies, renal biopsies, etc.
Such an improvement is not trivial and should never be viewed as such even as the development of fully automated machine diagnosis progresses. It’s a natural evolution, especially from a regulatory and practitioner acceptance viewpoint. Practising surgical pathologists’ workdays could be made more efficient and less tiring with digital slide AI-based prescreening and improved interfaces. But for now, the human will still make the final diagnosis.
In the future, the device may be better than the pathologist, but to get there will still take time. Kind of like waiting for the home run spectacular, but a single will do just fine in the meantime.
David Wilbur is a Chief Medical Scientist at Corista. In addition he is a Professor of Pathology at Harvard Medical School and specialises in endocervical and endometrial glandular lesions in gynecologic cytopathology, computational pathology and translational molecular medicine.
We will be returning to London for the Digital Pathology Congress: Europe. Click here for more information.
Leave a Reply