WILL AI REPLACE RADIOLOGISTS ANYTIME SOON?

Artificial Intelligence (AI) applications are entering medical practice at a rapid rate and are expected to play a huge role in transforming medical care. AI particularly excels in pattern recognition from well-defined input with binary outputs. This is particularly relevant in the field of radiology. In fact, Prof Geoffrey Hinton, the godfather of neural networks, said in 2016 that it’s “quite obvious that we should stop training radiologists” because of the trend toward using AI in radiology.

Just a few weeks ago, Google AI system surpassed a team of six radiologists in identifying breast cancers from mammograms, reducing the number of false-positive and false-negative results, but… the system also missed some cases, that all six of the radiologists had flagged as cancer![1]

We have indeed, seen some major performance breakthroughs recently, and radiology is a prime candidate for early adoption of AI….In 2018, the FDA approved the first algorithm that can make a medical decision without the need for a physician to look at the image.[2] The FDA did not stop there, and currently continue to approve around one medical imaging algorithm/ month.[3] Tech is moving fast for sure, but its integration into every day clinical use seems to be a different challenge.

In this article, I will refer to AI as programming systems that perform tasks which usually require human intelligence (AI) - But also for algorithms that are trained to solve tasks using pattern recognition or with deep neural networks (Machine Learning and Deep Learning).

CHANGE IS A CONSTANT !

Radiologists have been there before. First, in the 70s, some thought that the introduction of MRI Technology was going to replace radiologists. Physicians would immediately know everything they needed without a need for further radiologist’s explanation. This was not the case. In fact, we now need radiologists more than ever, given the complexity of findings in MRI.

Then, the introduction of Computer-aided detection (CAD) systems in the 1990s. Based on pattern recognition software, they identify suspicious features on the image (i.e. for breast cancer detection in mammograms). This “revolutionary” technology, also known as “ AI-lite system” was supposed to decrease false-negative readings, but it was more time-consuming and difficult to use, given the higher rate of false-positive findings. Research suggests that they didn’t necessarily make a big impact on radiologist accuracy one way or another.[4]

CAD didn’t make radiologists irrelevant. They are simply new tools in the radiologist’s toolbox. These tools are required for radiologists to handle a lot more data quickly, make more diagnoses, and hopefully with better accuracy. This alone, dramatically oversimplifies what radiologists do…

Let’s take a CT scan for example. It involves eight fundamental technical parameters that can be altered or optimized to lower the radiation dose to the patient, while preserving the best diagnostic image quality.[5] This inherent trade-off, tailored to the patient’s condition, could be difficult to be appreciated by radiologist[6] and even more by machines (for now…

The real question is - “What is the real ‘Job-To-Be-Done’ of radiologists?” Is it just about looking at white spots on an X-Ray?

They look for numerous conditions all at once, while also keeping an eye out for anything else suspicious, taking into consideration medical records and other lab results, for example, patient’s risk factors and clinical presentation, etc. They also treat diseases, interact with a wide range of referring physicians; neurologists, urologists, orthopaedic practitioners, etc, for diagnosis and treatment, talk to patients, and serve on multi-disciplinary boards, teaching, training, etc. The interventional radiologist[7] delivers appropriate imaging-based therapies.

Let’s have a look to a typical day of a radiologist, and where AI may impact.

WHERE AI CAN HELP: EFFICIENCY AND WORKFLOW OPTIMIZATION

AI, many believe, can and will, continue to optimize radiologists' workflows, which is probably the most immediate, and obvious benefit. “Radiologists can look at 360 slides of CT scan. But with the technology becoming more sophisticated, how to look through 20,000 slides? It’s the same for genetics. You have a terabyte worth of data in your microbiome. It crosses the threshold of human capability” mentioned Dr Jeremy Lim, medical doctor, entrepreneur and Partner in Oliver Wyman’s Singapore.

Radiology AI systems perform single tasks (narrow AI) which could be considered as monotonous, for radiologists, (i.e. quantification of specific distances, making measurement of heart function, delineating the cardiac chambers on cardiac CT, etc.). They could also perform repetitive and tedious tasks, such as appointment booking, counting and measuring metastases, or enhance the diagnostic precision of such tasks. Rad AI, a US based start-up, has been focusing on automatically generating the impressions section of a radiologist’s report – This is currently saving radiologists 60+ min/ day.

WHERE AI CAN HELP: FINDING COMPLEX AND SUBTLE PATTERNS

Beyond the efficiency factor that AI brings, the “image problem” is another area that recent developments have proven quite impressive. AI algorithms have shown that they can label data better than humans do, meaning it can effectively detect and diagnose from images, things such as bone fractures, potentially cancerous lesions or haemorrhage on the brain with MRI.

In this picture below, on the left it’s what 90% of doctors see using CT scans, and on the right, the potential problem areas, identified by Brainomix’s technology. It provides rapid and standardized assessment of stroke patient’s CT scans.

Arterys and Aidoc are two series C start-ups, known to identify patterns in images with high reliability, and find complex and subtle patterns. Their software can deliver results that are mostly similar to what radiologists would otherwise do manually (or better).

Another interesting example, of trying democratizing echocardiography, is Eko.ai, a Singapore based startup. Their machine learning platform automates the analysis of echocardiograms to make diagnosis faster and more accessible to patients. Currently, measuring and interpreting an echocardiogram requires 30 minutes of a cardiologist’s time, and 250 clicks. Further-more the process of this being done manually, can also result in up to 20% variability…Dr Carolyn Lam, a renowned cardiologist, researcher, and also co-founder of Eko.ai, wants to bring this process to “a 2-min process, a single click and 0% variability. Our ultimate goal is to put heart health screening into everyone’s hands.”  This not only speeds up the diagnosis of heart disease by increasing access (and accuracy) of diagnosis (plus access to treatment), but also increases the overall efficiency of healthcare professionals and systems.

Dr Lam also mentioned ongoing partnerships with pharma’s to improve the performance of cardiovascular clinical trials, which “ultimately is about moving towards precision medicine, simply having the right drug, the right dosage, for each patient at the right time”. Something that I am also quite excited about!

So, yes, the real power of AI today, and the near future, is pattern recognition in imaging, and its ability to identify and diagnose, with high reliability, that is often more complex than humans can (.ie. nodule on chest CT or haemorrhage on brain MRI). For the most part though, ‘the best systems are currently on par’ with human performance and are used only in research settings.

Current models are only trained for specific tasks in image recognition, the issue being, thousands of these narrow detection tasks are necessary to fully identify all the potential findings in medical images, yet only a few of these can be done by AI today.

“Let’s take an example in ophthalmology, looking at Diabetic Retinal Photography. About 0.5m Singaporeans are screened every year, so about a million of eye pictures to be analysed... 80% are typically being classified as normal. So, what if AI can eliminate this 80% and help our healthcare workforce focus on the remaining 20%? It will be a massive productivity gain that will also translated to clinical outcomes improvement for our patients” mentioned Dr Lim.

WHAT IS NEEDED FOR AI TO LEAD IN RADIOLOGY…

…and in a broader perspective in healthcare? If AI applications may solve the dilemma of what's known as the “iron triangle” in healthcare, in which three interlocking factors—access, affordability, and effectiveness—require inevitable and often negative trade-offs, it will also need to address the question of accountability and stakeholders’ adoption:

• ACCESSIBILITY – In remote areas or under privileged areas where indeed AI may be most needed. Population simply don’t have access to the same technology, resources as in the US or in EU, or have simply not enough radiologists to start with. Zebra Medical Vision decided to address this problem, in offering its algorithms to healthcare providers globally for only $1/ scan. Last year, they announced a partnership with Apollo Hospitals (70 hospitals and 200 care and diagnostic clinics) in India to focus on the development of an AI-based chest X-ray interpretation tool for the early detection of tuberculosis;
• AFFORDABILITY (and/or effectiveness), which will be about reducing the costs and increasing the efficiency of healthcare processes in general. AI should serve to augment and expand a radiologists’ job description. It will optimize the overall workflow, make healthcare providers more productive, give flexibility for spending time with the more complex patients, the ability to make better and more accurate diagnosis, and allow them to allocate their time differently. Everyone is looking at AI as a “doctor augmentation tool” to augment the care provided;
• ACCOUNTABILITY – some form of safety net needed and the question of “data security, trust and the black box reasons. Who is to blame if an algorithm gets a diagnosis wrong? Clinicians (and patients) need to know the reasoning behind a particular decision, which is sometimes known as the interpretability problem…” mentioned Dr Lim;
• ADOPTION by multiple stakeholders, including the “users”, “influencers”, and “purchasers”– doctors, other HCPs, governments and most of all, patients/ public may hesitate to embrace AI which does not offer the human touch/ faceless technology.

Dr Lam also raised the importance of, not only having the right algorithm, but also having the right solution or “commercial product” that can be used everyday, in clinical use. “It’s not because you have an algorithm that you have a product…We should not underestimate the amount of work that it takes to integrate with existing EMRs, fix data silos, design the right interface, protect patient data, get all the regulatory approvals, etc.”

WHAT’S NEXT FOR RADIOLOGISTS?

Today, one-third of radiologists’ cases are cancer-related, where histopathology ultimately is the standard reference for definitive diagnosis, leaving another third related to infection (e.g. COVID-19) and the remaining to trauma, auto-immune diseases, etc.

”With a more evidence-based approach facilitated by big data and analytics, augmented with the precision of AI in computer vision, we may see even greater accuracy of radiological diagnoses for specific conditions that approach the reference standard of histopathology using imaging as a non-invasive means to diagnosis” mentioned Dr Cher Heng Tan, Assistant Chairman of the Medical Board for Clinical Research and Innovation,  a radiologist at Tan Tock Seng Hospital in Singapore – In other words, the ultimate goal for imaging, augmented by AI, will be how histopathology or microbiology is today, as a reference standard for diagnosis.

Let’s also not forget that AI algorithms are often compared to radiologists, based on their ability to identify a single disease, which as we have seen, is not the case. AI will replace radiologists as pure interpreters of medical imaging in the mundane tasks such as measurement and segmentation. But, AI cannot replace the human brain behind the radiologist. Perhaps what would be more likely to replace radiologists, would be radiologists who embrace and are augmented by AI.

Finally, radiologists will need to apply a “disrupt from within" approach, lead the change of their own practice, be the early adopters of AI technology and be ready, to be the future leaders in healthcare AI technology. A new profile of radiologists will therefore emerge, adept at medicine and computer science. After all, Microsoft Excel didn’t replace any accountants, nor did autopilot systems in aviation replace human pilots. AI is no different. We are only at Chapter 1 today, and there will be much more innovation and development to come in Radiology with AI.

[1]  International evaluation of an AI system for breast cancer screening, Nature, 2020

[2]  Rise of Robot Radiologists, Nature, 2019

[3]  Eric Topol, author of “Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again”, 2019

[4]  Will Artificial Intelligence Replace Radiologists, Curtis P. Langlotz, RSNA, May 2019

[5] CT scan parameters and radiation dose: practical advice for radiologists”. Raman SP1, Mahesh M, Blasko RV, Fishman EK. NCBI, 2013

[6]   A questionnaire survey reviewing radiologists’ and clinical specialist radiographers’ knowledge of CT exposure parameters, S. J. Foley, M. G. Evanoff & L. A. Rainford, Springler, 2013

[7] Interventional radiologists, not only interpret medical images, but they also perform minimally invasive surgical procedures through small incisions in the body.