By Steve Potts, PhD
The formal training of pathologists has an impressive history, and pathology boards for both medical and veterinary pathologists are very difficult professional exams. As digital pathology slowly makes it way from a early adopter technology to mainstream, there will be an increased need for training of the “digital” pathologists, both in working remotely with digital slides (which is fairly quick to pick up) to image analysis (which takes more time and specialization). The vendors of digital pathology scanners generally offer basic and generic software training, and have done a good job of educating the pathologists in how to use the tools. However, much more work needs to be done across the industry in improving training opportunities for pathologists using quantitative analysis, both in the research and clinical markets.
The challenge of image analysis training in the clinical environment is illustrated by a question I had from the audience at a talk I gave recently on Validation of Image Analysis in GLP and CLIA environments. Someone asked,
We use the ER/PR algorithms from the manufacturer on Ki67 out of the box, and we want to know how to validate them?
The problem with this simplistic approach is that Ki-67 has a binary thresholding criteria rather than multilevel in ER/PR, and the Ki-67 staining patterns are generally much different in the nuclei than ER/PR and usually require tuning changes in how the software scores. And Ki-67 is a very simple marker! The best validation process is useless unless the pathologist knows how the algorithms work, and has a generally good understanding of image analysis concepts. While we are working actively on better validation approaches for image analysis for both clinical and GLP preclinical environments, real world training in image analysis must come first.
So prior to our plans for image analysis validation services, we first need to make training of image analysis more accessible to pathologists as well as imaging scientists and support personnel at pathology centers. This is a difficult activity for vendors to do, as it really should be taught by a combination of pathologists and image analysis experts. So many image analysis results heavily depend on the condition of the tissue in staining, fixation, or other histology steps, the pathologist needs to be involved in training to be able to point out some of these things. IHC overstaining is the precise reason that pathologists need to work closely with histotechnologists in determining the exact antibody dilution to use with each stain. Close communication avoids other artifact pitfalls that affect pathology endpoints and analytical consistencies. Some antibody vendors actually customize their IHC processes to make the antibody stains much darker, which might(?) be helpful to the human eye, but is poor for image analysis, where lighter staining is usually much better. If I had a dollar for every time someone has asked for an algorithm that will work on overstained IHC slides, well, I could buy a couple more slide scanners.
The other reason it is difficult for vendors is that under strict medical device manufacturing guidelines, particularly in the clinical environment, vendors need to stick to discussing and training on the markers for which they are cleared. This means training on the familiar three amigos ER/PR/HER2, and usually on only the subtype of breast cancer for which they have been validated and cleared. There is much more to this limitation than can be addressed here, but it clearly limits for good reason clinical IHC software vendors conducting broad training in image analysis.
The other challenge in training pathologists is time. Due to the heavy workload of many pathologists, there is often little time to devote to attendance at a two-day course either onsite or at the scanner vendor location. The histologist and the lab manager may go, but usually the pathologist simply cannot get the time free in his or her schedule.
To respond to this need, Flagship Biosciences is offering a once-a-week, six week image analysis course online, scheduled each Thursday. It is co-led by an image analysis expert and a board-certified pathologist, which makes for a fascinating interaction. Each session only lasts 2 hours (we may cut it to 90 minutes in the future, this is the limit that people can absorb information), and it gives a week in between for attendees to try out the concepts, see what worked and didn’t work, and come back with questions. We will start with basic cell and area measurement, and move toward more advanced, sophisticated software-aided approaches as the courses progress. We hope that after a pathologist and a pathology lab knows how the algorithms work and how to relook at their IHC approaches for quantitation, the next steps will be image analysis validation training.
It will be interesting to see if image analysis becomes something every pathologist is familiar with and can run, or will be the domain of pathology specialists. I hope for the former, but I am betting on the latter.
