Imaging Informatics: Breaking Down Siloes
By Ittai Dayan, MD
Vol. 24 No. 4 P. 6
Federated learning models help disperse data.
Over the past few years, radiology has become the leading specialty in adopting AI solutions in health care. Radiologists increasingly rely on AI tools to diagnose and treat patients, with some estimates of clinical adoption by radiologists estimated at 30%, as of 2021. In addition to supporting the clinical workflow, radiology departments also leverage AI solutions for automating administrative tasks, such as revenue cycle management and improving operations. In addition, radiologists are at the forefront of developing new algorithms for quality assurance, standardization of practice, and novel diagnostics, including those that use radiomics (referring to high-throughput extraction of quantitative features from medical images using advanced computational techniques). In fact, the ACR estimates that there are hundreds of “homegrown” AI algorithms already being used “in the wild.”
It is not surprising that radiology has taken this leadership role. Previous investments in digitization (ie, rolling out PACS and different workflow enhancement solutions) have not only generated massive amounts of data but also a workflow that is (relatively) scaled up and AI ready. This sets radiology apart from other medical specialties that have only recently adopted EHRs and are still grappling with the adoption of workflow enhancement solutions. This was driven, at least partially, by the fact that a digitized and optimized workflow actually cuts costs for radiology departments and now provides compounding benefits. Despite radiology’s leadership position, there are challenges to fulfilling the potential of AI, which novel techniques such as federated learning can help overcome.
There are clear challenges in taking the practice of radiology to the next level by broad dissemination and reaping the benefits of using AI solutions. Radiology departments and networks are now required to both assess and then integrate a record number of new software into their workflow and are often lacking the resources to do both. When deployed to production, many AI solutions do not live up to expectations. Research on large amounts of medical imaging data is facing challenges due to technology costs, operational difficulties, and privacy concerns. There is much innovation happening, but even at the current rate of software-as-a-medical-device going to market, only a fraction of the potential use cases in radiology are covered, which raises the question of AI’s ability to make a dent and deliver meaningful outcomes to a large number of patients. This is especially important given the high ticket price that organizations generally need to pay in order to enjoy the benefits of AI solutions. In turn, this presents the question of return on investment. Some organizations have managed to capitalize on the “AI gold rush,” while others are left out, often the organizations that would contribute diverse data that are reflective of the broad patient population (and often the patient population that needs AI the most).
A Different Model
Federated learning has the potential to change all of that. Federated learning is a machine learning technique that enables the training of models across multiple decentralized devices or servers without the need to transfer data to a centralized location. Instead of sending raw data to a central server for processing, it allows models to be trained locally on individual devices, with updates to the model being shared across the network. By enabling collaborations and alleviating the need to share data between multiple institutions while preserving patient data privacy and data security, federated learning can help to overcome some of the major challenges associated with data sharing in health care. The result would be a massive scale-up of radiology data, which are currently siloed in multiple PACS systems and data repositories. It will also enable the leverage of additional clinical data, which are critical in developing and assessing the performance of AI solutions.
Hospital-based providers and researchers have been at the forefront of federated learning for the last three years. Notable efforts include the EXAM study,1 which included 20 institutions worldwide in an effort to create a tool for predicting patient outcomes in the COVID-19 pandemic and, more recently, a study to generate an automatic tumor boundary detector for glioblastoma2 with over 70 institutions worldwide. Both of these studies leverage medical imaging and were published in nature group papers (full disclosure—I was the first author of the former) and demonstrated a new world of opportunities for federated learning enabled collaborations. By enabling an unprecedented speed and scale of collaborations, they are paving the way for hospital-driven research that knows no boundaries “data without borders,” as recently defined by the chair of the ACR commission on informatics, Christoph Wald, MD, PhD. Additional specialties, like pathology, are not falling behind, with a recent study published in Nature Medicine in January on predicting histological response to neoadjuvant chemotherapy in triple negative breast cancer.3
The results of the dissemination of this technology will be widespread. First, it will drive the dissemination and translation of research. AI algorithms that currently reside in one institution can quickly be adapted, validated, and tested elsewhere, and cross-institutional research efforts can scale without driving massive costs. Second, it will drive a paradigm shift in industry-provider partnerships, one that reduces the costs of collaboration, maintains security and privacy of data, and ultimately democratizes the field (a worn-out expression that underscores the existing gap in the market). Third, it will enable the development of AI in more use cases than can currently be afforded, with the study by Pati et al referenced above as an example for why accessing and curating data across multiple institutions is required for enabling research into rare diseases. Lastly, it will pave the way to continuously evolving AI solutions that can be useful in the radiology workflow without running the risk of being biased to a single data silo. While skeptics mention the difficulty in bringing together so many institutions and managing distributed datasets across the world, I am a staunch believer in the ability of the medical-scientific community to rise to the challenge.
— Ittai Dayan, MD, is the cofounder and CEO of Rhino Health.
1. Dayan I, Roth HR, Zhong A, et al. Federated learning for predicting clinical outcomes in patients with COVID-19. Nat Med. 2021;27(10):1735-1743.
2. Pati S, Baid U, Edwards B, et al. 2022. Federated learning enables big data for rare cancer boundary detection. Nat Commun. 2022;13(1):7346.
3. Ogier du Terrail J, Leopold A, Joly C, et al. Federated learning for predicting histological response to neoadjuvant chemotherapy in triple-negative breast cancer. Nat Med. 2023;29(1):135-146.