How Can we Develop Collaborative Standards for AI in Digital Pathology?
Posted 22nd March 2019 by Joshua Sewell
As laboratories transform their workflows into the digital environment, a tremendous opportunity presents itself: to transition the field of pathology from a qualitative to quantitative discipline. Quantitation brings measures of accuracy, reproducibility, and statistical stringency that allow computational algorithms (including AI) to perform complex tasks and measure their success. The evolution of Pathology will not be dictated by any single organization but rather will take an entire community of experts.
In this blog, I will highlight some of the opportunities for standardization in computational pathology. The goal is to educate others on existing frameworks and design patterns that will allow reusable code in any lab. Precedents for standards already exist in medically-focused Big Data domains of Radiology and Genomics.
In 1993, Radiology adopted the DICOM standard as a way for modular applications to transmit and receive data from medical images. The explosion of genomic sequencing based tests drove Molecular Pathology to accept best practices in the genomic sequencing field – such as the FASTQ, BAM, BED, and VCF file formats. The foundations provided by a common file format have allowed innumerable numbers of complex workflows to be constructed for ever increasing analytical demands. Digital pathology should be no different.
So what about AI? How can we build code that leverages commonalities across institutions and use cases? The first step to start addressing these key questions is to communicate shared problems and ideas.
For example, when building an AI algorithm to differentiate normal from tumor tissue, one usually requires smaller image “patches” to be extracted from a whole slide image. In the dozens of papers I have written, reviewed, or read, it has become increasingly clear that the small developer community is building their own custom solution – each essentially doing the same thing, but in different ways. How many hours have been lost because someone wrote a piece of code from scratch, rather than using an existing library?
There are even more questions: Could we all agree that one or two ways are the best practice on terms of efficiency and accuracy? What about standardizing the process for developing and validating an AI model? What is the accepted approach for training and assessing a model? What are the key criteria and minimum amount of information about an experiment that we should accept in our journals? Do we, as a community of software developers in the Digital Pathology area, have an expectation that all code and models be open source and in the public domain?
I don’t have answers to many of these questions; I can only speak to my lab’s experience. However, I would like to see the building of a strong informatics community that develops and supports the adoption and implementation of AI into Digital Pathology. I envision this as a collaborative effort – including the sharing and hardening of code – across both academic and commercial laboratories. Sharing and reusing modular code will allow us to prototype with less boilerplate, catch bugs earlier, and disseminate best practices as people come from other scientific disciplines. Defining standards for interoperability will be a boon to the development of highly sophisticated software applications, but only if we all work together.
Steven Hart is Associate Consultant and Assistant Professor of Biomedical Informatics at Mayo College of Medicine. He led a roundtable discussion at the 5th Digital Pathology & AI Congress: USA.
Don’t miss the chance to download the 6th Digital Pathology & AI Congress agenda for a full synopsis of topics under discussion at the congress.
Leave a Reply