Time for AI to Deliver

What is the role of AI in accelerating research? During a webinar in September 2020, Florent Saint-Clair of Dicom Systems and neuroradiologist Dr. Peter Chang of Avicenna.ai, discussed the growing role of machine learning in enterprise imaging.

COVID-19 Impact On Radiology

Dr. Chang shared his experiences “in the trenches” of radiology, as a practicing neuroradiologist at UC Irvine.
With the majority of non-urgent diagnostic procedures postponed in the early months of the pandemic, PACS are currently underutilized. Yet, despite the reduced volume of imaging studies at this time, radiology must remain ready for a spike in utilization. Once procedures return to their pre-pandemic or “new normal” schedules, the volume could jump on a moment’s notice. Imaging is expected to be there for patients, and 100% operational for when procedures come roaring back. Our focus at Dicom systems is to be the support mechanism for physicians and for medical imaging.
Dr. Chang shared that at the UC Irvine system, the current volume of imaging studies is approximately 63% of the traditional volume for this time of the year. Compared to other medical specialties that primarily serve one type of patient population, such as dermatology or gastroenterology, the impact on radiology has been less drastic.
Dr. Chang commented on the increased level of funding for imaging research as it relates to COVID-19. The National Institutes of Health have allocated $60M in funding for COVID-19 research; more locally, grants within the University of California system have enabled some of Dr. Chang’s recent work.

Dicom Systems Aims To Make Enterprise Imaging AI More Accessible

Since 2008, Dicom Systems has been a trusted partner for radiology practices and hospitals. With the advent of AI, our customers have approached us to play a key role in their adoption of AI. As our servers already function as the gateway for all their images, customers asked us to help them figure out how to adapt AI in the existing imaging environment. However, in order to unlock the value of AI to PACS/EHR, healthcare organizations need to overcome a number of obstacles.
Florent Saint-Clair explained that the largest barrier for AI is access to the cloud. While the best way for AI companies to operate is to host their solutions in the cloud, not all customers are able and/or willing to transition away from on-prem infrastructure. Due to procedures, regulations, and the need to avoid PHI related vulnerabilities,several of our large customers want to adopt AI, but don’t have the liberty to do so. This makes Dicom Systems a natural partner for their AI deployment.
To that end, we launched AI Conductor: a standards-based workflow orchestrator to enable our hospital customers to adopt AI algorithms. With AI Conductor, images come into our server (cloud or on-prem), and go through a filtering mechanism. If the images are flagged for further processing, they become a candidate for AI algorithms. Depending on the filter type, the images are shared with a specific docker via DICOMweb, each of which serves a different mission.
We are able to use the DICOMweb RESTful API to send data to the algorithms and to receive the results from the algorithm.Some results may be DICOM overlays  that can then be merged to a DICOM study. Others, thanks to our partnership with Arterys, leverage the dictation system to pre-populate diagnostic reports with results of AI. There are various ways to consume AI: as long as the solutions conform to industry standards, AI Conductor will process them and deliver them in the appropriate format.
The ultimate goal is to make AI more accessible to enterprise imaging. Today, many of the AI projects may not be clinically relevant right away. An example is an algorithm developed by a Ph.D. student. At the same time, many of the clinically relevant AI algorithms are ready to find a home to be “adopted” by our healthcare customers.
We encourage the industry to take a vendor-neutral and industry standards-based approach to deliver AI. This level of integration is essential for the adoption and consumption of AI in enterprise imaging.

Opportunities For Deep Learning In Enterprise Imaging

The last portion of the webinar was dedicated to Q&A.

Can the algorithms, once implemented, learn from the patient population?

To date, the FDA has cleared or approved “locked” algorithms only. They define a “locked” algorithm as one  that provides the same result each time the same input is applied to it and does not change. There is currently no approval for “learning” algorithms. These are described by the FDA as adaptive algorithms, for which the current regulatory frameworks were not designed. However, Dr. Chang, a panelist at a recent FDA workshop, shared that the FDA is now aware that continuous learning is one of the primary benefits of AI, and believes FDA approval for adaptive algorithms is not too far off.

Can AI be available to other disciplines, besides imaging? Can imaging and non-imaging data be combined in a deep learning algorithm?

While radiology has a head start in AI due to the long history of storing studies digitally, other disciplines, such as cardiology or pathology, can absolutely leverage AI. AI is also expected to give ophthalmologists new automated tools for diagnosing and treating ocular diseases, such as diabetic retinopathy (DR) and diabetic macular edema.

In fact, Florent Saint-Clair emphasized, AI must integrate all aspects of diagnostics in order to become more and more intelligent over time. The key to that level of integration is all vendors adhering to industry standards –DICOMweb Dicomweb and FHIR – to marry those results. In time, cross-domain standards will emerge and provide ways to integrate data from different sources and disciplines.

Can AI help thwart a pandemic by identifying animal viruses?

Scientists estimate that 1.67 million viruses exist on earth. A fraction of the 1M+ animal viruses have the potential to “jump species,” a process called zoonotic spillover. Whether or not a virus can make such a leap and cause the next pandemic depends on a complex combination of factors related to ecology, viral evolution, and human immunity.

Dr. Chang explained how AI applications could address this with “virtual experiments” in cases where we have a lot of historic data, but the cost of lab experiments is prohibitively high. An algorithm could be trained to predict which virus has the capacity to become a threat.

The COVID-19 pandemic has, without a question, upended healthcare worldwide. But not all changes it has brought about are disruptive or catastrophic. With increased funding, policy changes, and accelerated approvals, AI adoption in healthcare may become a silver lining of the pandemic.