How AI is enhancing patient care and improving radiologists' lives

When the topic of the future of radiology comes up in conversation, I frequently find myself being asked a few recurring questions: “How will artificial intelligence affect radiology?” quickly followed by “Do you think AI will replace you?”

These are great questions that I have found myself contemplating as well. As someone with an overall technological curiosity and a radiologist who embraces technological advancements, I ponder these questions often.

Let’s start with the first question.

How will artificial intelligence (AI) affect radiology?

The short answer is that AI will have a profound effect on many different facets of radiology, ultimately improving our accuracy, efficiency, and communication.

Here is how I foresee AI affecting and influencing aspects of radiology over the years to come, broken down into different parts of radiology — from image acquisition to worklist automation and image interpretation.

Image acquisition

Like many things in life, the majority of time spent on an imaging exam is in preparation — informing the patient about the exam, having the patient change, placing an IV (when necessary), positioning the patient, setting up the scanner, etc.

Some of these tasks, such as placing an IV, are not going away any time soon. Some tasks can be automated, e.g., patients can review and fill out forms/upload identification and insurance cards electronically, and pre-procedure instructions and directions to a changing room or procedure room can be administered electronically.

For more complex exams such as CT and MRI, AI will likely be able to help position patients appropriately and set up imaging fields of view while technologists work on other tasks. Some newer software packages are already auto-create and auto-send image reformats (sagittal, coronal, MIPS, etc.) based on the selected protocol, freeing up small amounts of time for technologists.

Fortunately, small things add up over time. If you can shave 5 minutes off each exam, you can scan a few extra patients per day. Given that many imaging centers already have growing backlogs, fitting in a few additional patients a day can do wonders for patient access.

Image post-processing

We discussed how some software packages already exist that will auto-process CT reformats. Another area where AI is poised to make an impact is the post-processing of MRI exams.

With MRI, some sequences are more time intensive than others, with some sequences taking several minutes to acquire enough data to create quality images. And where there’s a will, there’s a way!

Current MRI vendors and several new tech start-up companies are actively tackling this problem; several already have solutions ready for clinical use. These cutting-edge algorithms can generate high-quality images by extrapolating from smaller datasets, allowing for shorter scan times and potentially decreasing motion artifacts.

Workflow and worklist enhancements

The lowest-hanging fruit in radiology is workflow optimization. There is significant variability among groups regarding worklist management, ranging from a single worklist on a single Picture Archiving and Communication System (PACS, i.e., our workstations) to multiple worklists on multiple PACS across multiple health care systems.

While some basic worklist organization is possible with most out-of-the-box worklists, radiologists still spend time looking for the next appropriate exam to read. Which exam is closest to missing its turn-around-time (TAT) metric (factoring in class – outpatient, inpatient, emergency room/urgent care — and exam urgency — routine, ASAP, STAT)? And, with large, highly subspecialized groups, which exam is within the radiologists’ subspecialty/comfort zone?

Enter AI. With AI solutions such as Clario SmartWorklist, this can become automated with little thought required. Better yet, worklist management software such as this does a better job and performs more consistently than a radiologist (at least, this has been my personal experience).

Once you’ve selected and implemented the rules you want the software to follow, press “go,” and the software will feed you the next most appropriate case. And when you sign off a case, the next most appropriate case will automatically load. Now, I reserve my brainpower for case interpretation and other ancillary responsibilities (protocoling exams, fielding questions from referring clinicians, etc.).

Image review and interpretation

Image interpretation is the crux of what it means to be a diagnostic radiologist. We look at images, make key findings, and, with the help of clinical history, infer the significance of those findings.

Software solutions currently exist that allow for linked scrolling between current and prior exams. This streamlines follow-up exams by allowing radiologists to compare nodules and lesions more quickly, which is particularly important for cancer restaging and surveillance exams.

AI will be able to help with image interpretation through machine learning, with institutions like Stanford’s Center for Artificial Intelligence in Medicine & Imaging leading the way.

While far from perfect, basic computer-aided detection (CAD) software add-ons are available for clinical use in mammography and lung nodule detection. Deep learning algorithms are already learning from various imaging repositories and pathology databases and showing very promising results.

Future iterations of CAD will have the ability to make clinically significant findings, including relevant incidental findings such as abdominal aortic aneurysms, coronary artery calcifications, lung nodules, kidney stones, adrenal nodules, and much more.

Down the road, AI will likely be able to “screen” exams for critical findings such as central pulmonary emboli, pneumothorax, head bleeds, aortic dissections, free intraperitoneal gas, acute appendicitis, etc., and reprioritize these exams to the top of the worklist. This will expedite patient care, hopefully leading to improvements in patient outcomes.

AI will likely provide a “second set of eyes” on cases and will occasionally catch findings missed by the radiologist (unfortunately, we’re not perfect, despite our best efforts) or accidentally left out of the report (we’re frequently interrupted mid-case with clinical responsibilities).

AI will help radiologists overcome bias (satisfaction of search, anchoring bias, etc.) and improve radiologist accuracy.

Report creation

For radiologists, our final products are our reports. We combine relevant findings with the patient’s clinical history and synthesize our impression — what we think is going on with the patient. We organize our impressions by relevance, prioritizing the most clinically relevant findings.

We include clinically relevant incidental findings in our reports and make recommendations or suggestions to help guide the next steps in clinical management. We also occasionally recommend clinical correlation to help narrow down a differential diagnosis. When possible, we base recommendations on American College of Radiology (ACR) white papers composed of follow-up guidelines based on data and expert opinion.

In the future, this can easily be automated by AI tools, improving the accuracy and uniformity of follow-up recommendations between radiologists and across practices. This should result in fewer unnecessary tests, decreased medical imaging-related health care costs, reduced patient anxiety, and a higher level of patient care.

AI solutions, such as RadAI, also already exist that can read a radiology report and auto-generate an impression within seconds. While imperfect, software like this helps speed-up impression generation, decreases omission of clinically relevant findings from the report impression, and decreases voice recognition and typographical errors.

Communication of results

Communication is crucial in all aspects of life, and radiology is no exception.

As radiologists, we make clinically significant findings every day. We may even find multiple findings warranting follow-up on a single exam (e.g., I’ve seen up to four synchronous primary cancers on a single CT). Ensuring patients receive appropriate follow-up is critical — a patient falling through the cracks is one of my biggest fears as a radiologist.

AI to the rescue again! Patient databases with monitoring programs for indeterminate and incidental findings will likely become robust and help remind providers and patients alike of upcoming follow-up exams. Databases will also be able to update in real-time if or when follow-up is no longer indicated (e.g., an indeterminate adrenal nodule has since been characterized as a benign adenoma or a prior exam has established >2 years of stability for a solid lung nodule, both no longer requiring further evaluation or follow-up).

Will AI replace radiologists?

Predicting the future is impossible, especially when looking from the flat portion of the exponential curve. Scientific and technological advances will continue to move at breakneck speed. But will AI replace us?

Probably not within my career (I’m about seven years post-fellowship at the time of this writing). There are so many diseases that can present in so many different ways that we’re probably a long way off from AI being able to replace us. Even a radiologist nearing the end of a 30+ year career will share how they still see new pathologies and pathologic presentations all the time.

AI is unlikely to replace radiologists, at least in the near future, but radiology practices that embrace AI may end up replacing practices that do not.

Besides, software companies will want to avoid taking on the liability. Why risk a lawsuit when they can charge a time-based or case-based fee in perpetuity?

Final thoughts

Artificial intelligence is here to stay and will have a lasting effect on health care (as long as we can avoid Skynet).

AI will become an integral part of radiology. It will make radiologists more efficient, accurate, consistent, and timely. In essence, AI will make radiologists better, improve radiologist quality of life, and likely have a significantly positive impact on patient care. And with aging baby boomers, growing backlogs, and a worsening physician shortage with no end in sight, the timing couldn’t be better.

Brett Mollard is a radiologist.