AI Adoption in Radiology is following the Same Arc as Teleradiology: Skepticism → Slow Start → Exponential Value
By Arunima Rajan
Arjun Kalyanpur is the CEO and the Chief Radiologist at Teleradiology Solutions, Bangalore. In 2002, he founded Teleradiology Solutions, which currently provides teleradiology reporting services to over 100 hospitals in the United States, Asia and Africa. In an interview with Arunima Rajan, he says that the challenge in AI deployment today lies in the profusion of choices and the difficulty of knowing which solution to adopt.
You returned from the US and launched Teleradiology Solutions at a time when the concept was unfamiliar in India. What inspired you, and what were the initial hurdles you faced?
Teleradiology Solutions was started because I was unable to get a job after returning to India from the US! A chance conversation with the Chairman of Radiology at Yale led to a research project and proof of concept which culminated in my reporting for Yale from Bangalore, and subsequently incorporation of the company in 2002. TRS was started at a time even before Skype, when internet connectivity in India was slow, unreliable and expensive, and most of the early challenges were related to technology. Despite these hurdles, what was apparent from the very start was how effective Teleradiology was, especially when coupled with the international model where night could be converted into day, leading to better reporting accuracy, optimal radiologist time utilization, and thereby cost savings. The surprise in the voice of the Emergency physician in those early days, when they learned that I was calling from India to discuss a case is something that always stays with me.
Indian healthcare has seen rapid changes, with tech taking center stage. From your view, how has the teleradiology and healthcare AI market evolved since you started? What shifts stand out the most?
Teleradiology in India has had a long nascent period. Although we began offering services to Indian hospitals within the first few years of starting our work in the US, acceptance in India was initially slow, with a lack of awareness of its benefits, coupled with connectivity challenges and high costs. In the 2010s we began to see slow acceptance, although mostly in the NGO sector, where our charitable work through the Telerad Foundation was the main contributor to growth. Two changes then transformed the scenario, namely the sudden explosion in private healthcare leading to a relative and increasing radiologist shortage, and COVID, which legitimised the work from home/ remote reporting model. Both these trends have led to a dramatic growth in the teleradiology market in India, and also set the stage for incorporation of AI into radiology.
AI is often seen as the solution for faster diagnoses, broader reach, and improved outcomes. But in Indian healthcare, how much of this promise has come true? And where does it still fall short of expectations?
If I focus on my own specialty and Teleradiology in particular, I see many similarities between Teleradiology and AI, and therefore I see history repeating itself today with AI as it happened with Teleradiology. Both are internet/technology enabled solutions to crippling healthcare manpower shortages, both address quality and speed of diagnosis, both can function in a cloud environment, and both have been initially regarded with suspicion and concern. So having seen the past of Teleradiology up close, I feel I am in a position to predict the future of AI. The promise is real and genuine, although it is true that is still remains a promise, in that there are only certain specific use cases where it currently delivers value, ie. we see just the tip of the proverbial iceberg. However, just as the advent of GPU brought AI to the forefront, other technologic advances will continue to come that will shift the paradigms, lower deployment costs and serve as tipping points to accelerate the deployment of AI in healthcare across all specialties and practices. Teleradiology is itself one such enabler, in that it allows radiology AI to be deployed at a global scale and benefit both radiologists and patients globally, thereby simultaneously delivering global impact. Going forward large language models promise to transform the manner in which medical content such as radiology reports are created efficiently, thereby saving the precious time of radiologists.
Which AI tools have moved past pilot projects and become part of daily hospital life? Are there areas where the hype still races ahead of reality?
Again, focusing on Radiology, where I see AI making its greatest impact is a) in the emergency scenario, think stroke and trauma, where time is of the essence and b) in large public health projects where vast amounts of screening data is collected that needs considerable radiologist time to analyse, as with breast cancer or TB detection.
In stroke in particular, our work in teleradiology has shown that AI has been transformative in improving patient outcomes. Hospitals across the world have adopted AI technologies that analyse CT scans to detect and quantify stroke, differentiate hemorrhagic from ischemic stroke and identify vessel blockages. Our own group has developed an Algorithm NeuralAssist which can detect, quantify and localise hemorrhage in the setting of stroke and trauma, identify the thrombosed middle cerebral artery which forms the nidus of the stroke, detect brain swelling and skull fracture in a trauma patient. This algorithm is currently being deployed at NIMHANS and within our own Emergency Teleradiology practice.
At the other end of the spectrum, our MammoAssist Algorithm helps detect early breast cancer on Mammography, an imaging modality which is time consuming to interpret and where the risk of error can significantly impact on patient life expectancy. In large national screening programs for breast cancer where the number of studies is enormous, AI such as this can be transformative.
AI currently adds great value in the setting of triage. By identifying the positive study for example in stroke, an alert system can be activated by the AI algorithm that can ensure immediate physician response that can greatly benefit outcomes. And in the TB screening environment, the AI algorithm can immediately filter out normal studies allowing the radiologist to conserve time and energy for the analysis of the positive cases.
As we know, the larger the dataset the AI is trained on, the better its performance tends to be. As a result, in relatively rarer conditions, the output of AI may not meet the expectations.
The rural-urban healthcare gap is a recurring issue. Have you seen specific instances where technology has effectively improved care in remote or underserved areas?
Our experience deploying teleradiology in the state of Tripura through our PPP model has shown us how remote areas can benefit from tech. We deployed a statewide teleradiology solution at the level of district hospitals and CHCs using a digitisation workflow in locations which did not have computed radiography, to deliver high quality reporting services to remote tribal populations. This was a truly impactful and fulfilling project that bridged the rural -urban gap.
Our international teleradiology work also benefits patients in remote and rural parts of the globe who present to emergency rooms and acute care centers in the emergency setting. By diagnosing conditions ranging from pneumonia to a ruptured appendix or aneurysm, we are able to improve access to quality diagnosis across the globe.
New technology often brings along messy data, difficult integrations, and resistance to change. What are the main challenges hospitals encounter when adopting AI, and which practical solutions have proven effective?
Clinicians and radiologists in particular are overworked today and lack patience as a result. It is therefore imperative to ensure that their time is least impacted by the deployment of an AI solution. A seamless workflow integration which requires the minimum orientation time goes a long way towards building clinician support and trust.
From the perspective of algorithm performance, it is important to collect the feedback of the radiologist in a manner that does not disrupt their workflow. Having a feedback loop that communicates newer developments and performance enhancements is necessary as with any technology. Deep learning algorithms are constantly improving their performance and therefore the trust and confidence are constantly growing. A single example of a subtle intracranial hemorrhage missed by a radiologist and detected by AI can itself be transformative in building radiologist trust in an algorithm. A high number of false positives which can sometimes occur in the early stages of algorithm development, has the oppositeeffect. It is therefore important to have a high level of performance validated independently in a core lab environment before deploying an algorithm clinically.
A hospital CEO wants to dip a toe into AI without chasing shiny toys. Where should they invest time, money and attention to create lasting value?
The take home message would be for the CEO to identify his/her own most significant pain points in the clinical space and look for solutions that address those.
For example, an orthopedic hospital CEO would do well to evaluate the range of products that autoanalyzer bone xrays and spine MRIs which can greatly help alleviate his radiologist scheduling concerns. Similarly, a neuroscience center would do well to look into AI algorithms in stroke care.
Large general hospitals can benefit most from implementing across the board workflow solutions in areas such as emergency triage, report generation LLM etc that can overall improve productivity and lower costs and enhance the patient experience, while simultaneously increasing throughput.
In summary, AI does indeed have the potential to deliver tremendous lasting value in healthcare. Those of us intrepid enough to test the waters and be early adopters stand to gain, although as with the deployment of all new technologies, patience and persistence are the virtues that will yield maximum benefits.
Got a story that Healthcare Executive should dig into? Shoot it over to arunima.rajan@hosmac.com—no PR fluff, just solid leads.