We’ve all heard about the promise of artificial intelligence (AI) in pathology. From the efficiency and productivity gains that workflow applications can deliver to the possibility of improved patient outcomes from predictive and prognostic applications, the potential is clear.
And this potential is largely what we’ve heard about for the past several years. Until recently, much of the conversation about AI in pathology has focused on how it will transform the field – not its present impact. It’s true that research has focused on developing deep learning algorithms, but these models are often not designed to work in practice; they’re trained on only a small portion of the pathologic entities encountered and not generalizable to real-world laboratory settings. Research alone is not enough to implement algorithms into practice. From having a means to deploy them to ensuring pathologist support, many labs haven’t been ready to adopt AI.
Recently, however, this has all started to change. We’re now seeing more and more evidence indicating not only that AI is ready for the lab, but also that labs are ready for AI.
Among this growing body of evidence is the pathology industry’s most comprehensive validation study to date (1), whose findings reveal a deep learning system for dermatopathology that can classify whole-slide images of skin biopsies with up to 98 percent accuracy across multiple laboratories. Our team developed the algorithm using a large volume of data to account for the wide variety of morphologies seen in practice. We then tested the model using uncurated data from three different laboratories to ensure that it could handle different methods of biopsy, tissue preparation, staining, and digital scanning. Although these sources of variability are not difficult for a pathologist to overcome, they can easily impact the performance of an algorithm if it is not properly developed. Our model, for instance, uses an image adaptation process to compensate for stain and scanner variations, then selects image regions that represent pathologic features. These suspicious regions are used to predict a classification for the entire whole-slide image. Finally, the system assigns a confidence score to each of its predictions, so that classifications aren’t provided when they are likely to be incorrect.
This deep learning system signals an advancement beyond ad hoc algorithms that are trained and expected to perform consistently within only a single laboratory environment. It allows any lab to sort and triage cases based on the model’s predictions. More generally, our findings support the growing impact of AI in cancer diagnosis, serving as the foundation for faster and more accurate diagnosis to improve treatment decisions and patient outcomes.
The time is now for AI-enabled digital pathology. Why? The answer lies, to a large extent, in the acceleration of digital pathology adoption. Moving deep learning beyond research and into practice requires a means of deploying applications so that pathologists can use them in their day-to-day work. Digital pathology platforms, which labs are steadily implementing to manage whole-slide images, serve as a natural launchpad.
These platforms address a variety of use cases – from primary diagnosis to improved sharing and collaboration. Granada University Hospitals in Spain, for example, has had a full digital implementation for primary diagnosis since 2016. The University of California, San Francisco began using digital pathology to reduce the time it took to share and diagnose frozen sections among its three hospitals and is approaching fully digital operation. We’re now seeing a new wave of labs flock to digital pathology for remote use in light of the COVID-19 pandemic and recent FDA guidance aimed at expanding its availability.
Not only are a growing number of digital labs well-positioned to adopt AI, but pathologists are recognizing its value; according to a 2019 survey of 487 pathologists from 54 countries, 73.3 percent of respondents expressed either excitement or interest about adding AI to their diagnostic workflows and 71.7 percent felt that AI could increase or dramatically increase diagnostic efficiency (2). Although the full benefits are just starting to emerge, when it comes to AI-enabled digital pathology, it’s clear we’ve moved beyond the hype.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
References
- JD Ianni et al., “Tailored for real-world: a whole slide image classification system validated on uncurated multi-site data emulating the prospective pathology workload”, Sci Rep, 10, 3217 (2020). PMID: 32081856.
- S Sarwar et al., “Physician perspectives on integration of artificial intelligence into diagnostic pathology”, NPJ Digit Med, 2, 28 (2019). PMID: 31304375.
