What Digital Pathology can learn from Radiology
Posted 17th May 2019 by Joshua Sewell
Radiology is ahead of the curve because they’ve had CAD (Computer Aided Detection) for about 20 years. Radiology as a field has therefore had experience of introducing and integrating AI algorithms.
In my previous post, I talked about high-level cross imaging modalities. Here, I will discuss three challenges specific to pathology. I also work with Radiology imaging, and I think that comparisons between the two can help see how pathology might develop in the future.
In pathology, less than 5% of practices are fully digital, which is a small fraction of the total number of pathology practices. This comes back to the problem identified previously about how to build these models: we need a lot of data. There is so much information in pathology and therefore the tremendous potential for computerised analysis of digital pathology to provide these kinds of applications.
However, we’re not going to get the quantities of data necessary to build applications that can improve the assessment of digital pathology unless practices are digitising all their data.
In my understanding, most of the incentivisation for pathology practices to go digital is purely related to workflow, as practices have developed and consolidated across multiple sites. The incentive to go digital has to do with workflow and not with a computerized assessment of the images.
At the point where the majority of practices are digital in pathology (as they are in radiology), we’ll see more applications which innovate the precision of care and improve medical practice. From my perspective, digitisation in pathology is the first challenge to be dealt with.
Large Image Size
A challenge unique to pathology is the size of image files (orders of magnitude larger than radiology). When digitising entire collections of patient images, the ability to digitise and store these images in a cost-effective manner has been holding back the digital pathology field.
However, storage costs are dropping and so this may not be a major problem long term.
Integrating AI into clinical practice
How does the practitioner take AI results and record them? How do they track agreements and disagreements between the medic and the machine?
Radiology has dealt with these questions for a while, whereas pathology hasn’t yet developed its own answers. The American College of Radiology has created registries to study issues of quality. They now have a data science institute that’s going to be systematically studying how well AI algorithms are working.
Pathology would benefit from developing similar types of projects, but it’s behind in doing this because there are not enough practices that are digital yet, nor many new applications related to decision support. But as more practices become digital and more applications start appearing, I think pathology can learn a lot by looking at how radiologists adapted to CAD.
Conclusion – Starting with education
Education is key to developing in this area. To enable positive steps forward, the pathology community can start looking at trying to incorporate digital pathology education into either residencies or postgraduate courses.
This will give rising pathologists insight into the possibilities of how practice can be enhanced and how it is evolving as practices go digital. An educated community is critical to the appropriate use and development of these technologies.
Daniel Rubin is Professor of Biomedical Data Science and Radiology at Stanford University, USA.
Leave a Reply