AI: from Lust to Dust

Ladies and Gentlemen, this is your captain speaking. We will have to push back to the gate to check a fuel pump warning light; we expect only a minor delay after maintenance has checked it.

The minor delay ended up being several hours because after restarting the engine, the light still came on, so we had to deboard and wait for another plane to be available. This was not my first delay and won’t be my last; such is the joy of travel.

The good news is that I had plenty of opportunities to talk with the person sitting next to me, who was also traveling to the annual radiology event in Chicago, which the RSNA organized. He managed a mid-size radiology group (<100 people) in South Texas, and I asked him about using AI in his practice. He explained that his physicians have been very enthusiastic about deploying generative AI as part of their reporting and can gain a 20-30% increase in productivity. They are still skeptical About deploying AI for detection and triage and have not taken the jump yet. Even though this observation is anecdotal and not statistically relevant, it confirms my suspicion that we are not quite there yet.

My absolute highlight of the conference was Eric Topol's plenary lecture. He has been my hero since one of his first books, “The Creative Destruction of Medicine,” came out more than 10 years ago. He gives the AI revolution an entirely new perspective. His presentation influenced my observations during the annual RSNA event, and I’ll use some of his quotes as I go on.

Key observations were:

·?????? RSNA was back in full swing: Attendance was up by double digits, especially among the professionals, and all vendors I talked with were positive. It was often hard to get a hold of people at their booth, and most of my insightful discussions took place during social events and receptions. Radiologists are key to the decision-making process; unfortunately, as Eric Topol reported, 800,000 Americans die or are permanently disabled each year by diagnostic error. AI promises to make at least a dent in that statistic.

·?????? It was all about AI: Almost one-third of the South hall, aka the RSNA AI Showcase, was packed with 100+ AI vendors, with the AI demo area “Radiology Reimaged” (aka “connectathon”) being the first to hit when entering the showcase area. Putting this AI demo together is a big job; kudos to #Mohannad Hussain, who has done a great job. I recognized several PACS veteran professionals in the AI booths who jumped the corridor and took on marketing, sales, and product management positions in AI companies; there is a shift in the mature PASC market towards AI companies with regard to investment and people.

AI encompasses two aspects in radiology: what I consider the “front end,” i.e., any AI that uses images as a source and outputs detection, triage, and orchestration until it appears on a radiologist's worklist, and the backend, which is mostly generative AI and assists in creating a diagnosis and a report.

·?????? Back-end AI is being deployed widely, Front-end AI not so much: My initial impression based on my discussion during my delayed flight to Chicago was confirmed as the week went along, i.e., there are still many hurdles to jump about the front end. Interestingly, the reverse is true for Europe, where there is more acceptance for front-end detection and more resistance for back-end detection. Here is what Eric Topol had to say about the significant barriers as of today:

·?????? There is resistance to change. This is not specific to AI but to any new technology and workflow. People are comfortable in their “zone,” so why change?

·?????? Reimbursement is still an issue in the US. Only a few algorithms can be submitted for insurance coverage. Charging patients an out-of-pocket fee of $40 for a mammo X-ray AI is not unusual, but I would be worried about the ethical considerations, as not everyone can easily afford that amount.

·?????? Regulatory barriers still exist as the FDA has much to catch up on. Most of the 1000+ AI FDA approvals include algorithms that use very shaky data. Most importantly, the FDA focuses on approving single-modal vs multi-modal algorithms, the latter of which will be more impactful. An example of a multi-modal AI algorithm would be to combine reports/texts and images, and the combination of AI with genomics and proteomics (protein information) has been very promising.

·?????? The models need to be transparent. In the opening session, Curt Langlotz, the RSNA president, suggested using “AI model cards” like nutrition labels to help radiologists decide if an AI model will work in their practice.

·?????? Compelling evidence is not available yet. The improvements in accuracy achieved are marginal, although if you consider the 800,000 diagnosis errors mentioned above, a 1% improvement would impact the lives of 8,000 patients.

·?????? Trust is lacking. This is partly fueled by media reports and futurist predictions that radiologists will become obsolete.

·?????? Implementation has been cumbersome, and integration is poor, especially at the backend. For example, automatic parsing and mapping into diagnostic reports are not widely deployed yet, a “zero-click” implementation is not quite here, and last but not least, it typically takes 6-9 months for a new software application to pass all of the cybersecurity testing and vetting by the hospital security team.

·?????? AI facilitates Individualized Medicine from “Prewomb to Tomb,” or, as Eric Topol quoted from “Lust to Dust”: Before birth, AI can predict based on the parent's genomes combined with fetus imaging such as an ultrasound certain conditions so one can be prepared during the delivery and early years. Undiagnosed diseases can be predicted and treated at adolescence. Cancer, typically in middle age but emerging at earlier ages, can be predicted early on. Coronary Artery Disease can be prevented by identifying high-risk individuals and implementing active surveillance and preventive strategies. Imaging AI plays a critical role; for example, a chest X-ray can indicate the following opportunistic interpretations: Type 2 diabetes, your coronary calcium score, the ejection fraction, the 10-year risk of a heart attack, or your risk from cardiovascular disease.

·?????? The most essential impact of AI will be “The Gift of Time”:

·?????? Key board liberation: When you visit a healthcare practitioner, you most likely notice that the face-to-face time is drastically reduced and replaced with your doctor looking at their screen. AI combined with voice recognition can give that time back to the actual interaction

·?????? Synthesize patient data: Doctors complain that they must search through many documents, lab results, EKGs, reports, and EMR data before or during a visit. In addition, creating a discharge summary after an in-patient episode based on possibly 100+ individual notes, observations, and results stored in an EMR is time-consuming and can be easily replaced by AI

·?????? Screening review of all images: Many procedures are performed to screen for potential diseases, such as mammo X-rays for women over a certain age, CT lung scans for heavy smokers, chest X-rays for miners, and TB screening for immigrants. Many of these procedures have a negative finding, and AI can be applied to eliminate most of these procedures from the radiologist reading list. AI can replace a “second reader” for those cases where this is common, especially mammo procedures. A disadvantage is that incidental findings might be missed; if an AI algorithm is used to detect only TB, it might not detect a heart condition.

·?????? Routine, non-serious condition diagnosis can be performed without any physician intervention. AI can already diagnose diabetic retinopathy with full confidence and this application is FDA-approved. Examples of non-serious conditions could be a skin rash, which could be diagnosed and medication automatically dispensed.

·?????? Workflow streaming and prioritization: Triage is critically important for stroke patients, where “time is brain,” as indicated by a CT AI algorithm. Other life-threatening conditions, such as AAAs, can be prioritized in a physician worklist or created as an alert to be taken care of immediately.

·?????? Patient virtual assistant/coach: Reminders for medications, appointments, referrals, coaching a patient on healthy habits, exercise, and sleep, as well as providing a virtual assistant for simple routine questions, are valuable.

·?????? Where will the radiology imaging AI reside, and what is the value of an AI platform? The approach from PACS vendors with regard to AI deployment varies widely. As a practitioner interested in deploying AI, you should seriously discuss your PACS provider’s strategy. At the AI front end, some AI imaging applications reside in the modality, such as a breast cancer and density detection algorithm in a mammo unit. At the backend, an AI algorithm could be native to the reporting system. Alternatively, one could use a third-party or the PACS vendor’s platform or marketplace for AI applications tightly connected to PACS. Some of the PACS vendors white-label a third party (e.g., Philips), some of them have their own proprietary platform or marketplace (e.g., Sectra, Siemens), some have a restricted embedded mode (e.g., NovaRad), and some don’t even get into the AI business and depend on outside orchestrators and 3rd party platforms (e.g., Infinitt). Depending on your PACS vendor for AI deployment might be advantageous from a support, billing, and integration perspective. However, you must be aware that a tethered solution does not always provide best-of-breed and could limit your choices, in which case a 3rd party platform might be preferred. I was kind of surprised that many professionals I talked with were unaware of the existence of 3rd party AI platforms, it appears that those vendors still have a lot of education to do.

·?????? AI might be the death of contrast pharmaceutical companies. Remember Kodak? For a baby boomer, this is a trivial question as it was one of the most valuable brands in the late 1900s; for my grandkids, not so much, unlike Apple, which is currently number 1. Kodak disappeared and slid from among the world's most valuable brands into oblivion as they did not anticipate how digital photography would replace film. It might not be long before AI replaces or dramatically reduces the need for contrast media. Similarly to using AI to denoise an MRI image, therefore reducing the scan time by up to 30%, one can use AI to create images that appear as if contrast was applied. No wonder companies like Bayer are investing in AI and AI platform companies as it will be their ticket to survival.

In conclusion, this RSNA was busy again. AI was the number one topic, although it is not quite there yet, especially as we like to apply it from “Lust to Dust,” as Eric Topol, one of the great medical visionaries, will let us believe will happen in the not too far future. I enjoyed the event quite a bit, meeting old friends and colleagues and making new acquaintances in the meantime. Looking forward to the next event, don’t hesitate to comment and/or share your thoughts on this topic.

Herman Oosterwijk