5 Factors Potentially Slowing Down AI Adoption in Radiology

Radiology in the last 30 years has progressed even more rapidly with the advent of digitization and practices are constantly catching up.  These technology advances are good for patients,  and have the promise of streamlined workloads for physicians.  The latest in the radiology world is artificial intelligence. This technology has particularities that may hinder swift adoption:

  1. AI will replace the physician

AI, as with other types of algorithms, has been integrated into many industries in particular when a repetitive task is required to be done with high level of accuracy and can be learnt by watching examples.  This can easily lead to one of the most common misconceptions about AI, that it will replace physicians. Spreadsheets didn’t replace accountants! This simply is not realistic, the reason?  AI focuses on the repetitive part of the task; physicians take care of people.   AI is designed to work alongside physicians as a second pair of eyes.

It is important to realize that when a physician works with AI, they can outperform either one working by itself.  AI can identify missed findings and improve the experience for healthcare providers and patients.  On average  a general radiologist interprets one image every 3 to 4 seconds. This high throughput workload has been shown to  increase the chances of a missed cancer. In many instances, breast cancer screening not being an exception, images are now acquired in 3D and not only 2D, which leads to an increase in the number of images by 50 fold or more per exam.

How AI is helping: This second pair of eyes makes the physician more confident in decision making and therefore faster in interpreting.

  1. AI is difficult to integrate in existing systems

Any new system integration with the present PACS systems is definitely a challenge.  Many of the tasks outside of reading images such as interfacing with PACS, calling up previous studies, creating their reports etc. are all challenges for the radiologist.  Most PACS systems are monolithic and traditionally force vendors to create custom interfaces.  AI can use more recent deployment technologies such as cloud and web interfacing, which makes integration into existing infrastructures much easier. 

  1. AI carries security issues

Two aspects of AI raise potential security issues: security of the patient information if processed in the Cloud and explainability of the results. For instance, Therapixel addresses the former via strict cybersecurity certifications such as ISO27001 and de identification of Personal Health Information (PHI) whenever possible, then encryption before uploading to our cloud. The latter requires to dispel the fear of the ‘black box’ AI. First, the result of the AI must be delivered with enough details so the physician can easily verify the result in case of doubt. Second, in case of a discordant result, it should be possible to report the case and work with the AI vendor on the reason for the discordance. This is an opportunity for the AI to continue to improve over time.

To gain trust AI vendors need to demonstrate when and where the AI solutions are integrated into the clinical decision process.  Tools that are easy to use and make their job simpler should develop  trust that will increase productivity over time.

What steps can help: ACR (American College of Radiology)has only recently started releasing formalized use cases for how AI software tools can be reliably used. 

  1. AI versatility is limited

At present, AI’s performance is linked to its area of expertise.  This is a challenge as the AI learns from large amounts of data for that specific area.  Outside of medicine,  massive amounts of data are available to train the algorithms.  A good example of this is Netflix. It has tens of millions of data input for what movies or TV shows people like.   In radiology the number of training images and their associated confirmed diagnoses is more limited or not readily available.  Since data is limited, diversity (patient, equipment), or lack thereof, presents its own challenges and may create bias in the algorithm. That said, as much as the Netflix AI can’t predict what you’d like to read, a breast cancer AI can’t detect a broken bone and for now, this is a limitation of AI versatility.

  1. New financial paradigm 

When deploying an AI solution, facilities may prefer not to put patients’ private data on the cloud and instead invest in hardware for local on premise processing of the AI system. This decision must  be weighed carefully.  The cost of maintaining an AI-grade IT infrastructure and its inherent inflexibility of  with demand, as well as  the safe use of de-identified data on a Cloud that scales up and down automatically to minimize the operational costs.

Additionally, AI is evolving very rapidly and investing in an expensive perpetual license is daunting.  The Software as a Service (SaaS) model is perfectly suited for mitigating this uncertainty: no up-front cost, all-inclusive prices (forget about expensive maintenance contracts), only operational expenses (a real advantage for investment-starved post-Covid budgets) and typically only an annual commitment.

AI has the potential to speed up interpretation throughout the radiology workplace.  AI detects more cancers, improves workflows, reduces contrast usage and radiation dose, and shortens reporting times.  These are all ways of improving the user experience, by reducing sleepless nights and potentially  increasing revenue.

AI financial assistance: CMS (Centers for Medicare and Medicaid Services) is working on better financial reimbursement but for now the ROI (return on investment)will come from quality, efficiency and the value of a better patient experience:

AI in the future:  AI is still very young in this market. 

Despite the challenges, AI has opportunities to improve accuracy, decrease variability in interpreting breast images, and improve efficiency in delivery of care, especially as more studies are published”        Dr. Conant

Transform your practice for continued success by implementing AI in mammography. We want to walk you through the process. Book a free, 10-minute demo today to learn how you can bring another set of eyes to your practice for improved efficiencies in screening and patient care.