Current state of AI in radiology
Dr Vidur Mahajan, Associate Director, Mahajan Imaging, mentions that the day is not far when software algorithms would assist radiologists and physicians in making diagnoses and imparting treatment to patients
On 14th November, 2017, a team of Stanford University scientists led by Andrew Ng, considered the foremost machine learning expert in the world, published an open paper describing an algorithm that can diagnose pneumonia in chest X-Ray at a rate much better than independent radiologists at Stanford. While many questions have been raised on the methodology associated with the development and validation of the results of the algorithm, one thing is clear – healthcare is now definitely on the radar of artificial intelligence experts and the day is not far when software algorithms would be assisting radiologists and physicians in making diagnoses and imparting treatment to patients. The day of ‘replacement’ radiologists and physicians though, in my opinion is much further away.
The definition of artificial intelligence (AI) has changed through time based on variation of its applications, advancement in computing power and general dissipation of use-cases through society. In fact, John McCarthy, who coined the term AI in 1954 complained that as soon an algorithm or product ‘works’, it isn’t called AI any more. The reason for this is that AI is projected to be a future-state and hence, anything that works flawlessly now, is not given the glamour of being called an ‘AI product.’ Take Google Search for example, it might just be one of the most advanced AI applications around us, but no one really calls it that because it has become ingrained in our day-to-day existence. In truth, AI is the application of large scale statistical data analysis for predicting a result, having an in-built ability to learn from responses to such results. (For more insight into this, I would recommend reading www.waitbutwhy.com’s blog post on Artificial Intelligence.)
More specifically, in radiology, AI will bring about changes in two fairly obvious dimensions in the short to medium term – quality and efficiency.
Quality improvement in radiology using AI
The first, and maybe the most direct way that AI will improve the quality of radiology will be by acting as an ‘assistant’ to radiologists, improving the consistency and accuracy of radiologists’ reports – essentially making a junior or general radiologist as accurate as a senior or subspecialised radiologist. An example of this is the chest X-Ray algorithm being worked on by Qure.AI, an Indian radiology AI company. A chest X-Ray is one of the most basic, but also one of the most difficult radiology investigations for a radiologist to report – such an algorithm, that can ‘steer’ a radiologist in the right direction, as far as a diagnosis is concerned, can go a long way in getting to a correct diagnosis.
The other way quality will be improved is by the merging of the fields of AI and quantitative radiology (also called radiomics). This relatively old field of radiomics has gotten a push recently given the reduced cost of computing power today and also the great abundance of AI tools that can be integrated. An example of this is a brain volumetry algorithm developed by QUIBIM (Valencia, Spain), which enables radiologists and neurologists to calculate the volume of a patient’s brain and its various components, enabling objective evaluation of several neurological illnesses. This quantitative approach means that new findings can be extracted from existing data simply because computers are able to ‘see’ in many more dimensions than the human brain. In fact, more recently, studies have been conducted which blend radiological and genomic data (a new field called Radiogenomics) enabling prediction of outcomes of diseases by taking into account both, radiological and genomic data.
Improving efficiency of radiology departments and radiologists
While quality improvement in radiology might sound like a more distant concept, AI has already started bringing about huge improvements in efficiency in radiology departments. The most obvious application of AI in this sense is the reduction of time associated with certain complex analysis. A fairly well-developed example of this is a liver tumour segmentation algorithm made by Predible Health (Bengaluru, India) which cuts down on post-processing time associated segmenting out the liver vessels, parenchyma and tumour from 45-60 minutes to 5-10 minutes. Such algorithms are especially relevant in the developing world where companies have to operate in resource constraints.
The other way in which AI is bringing about efficiencies in the radiological world is by helping with triaging. Picture this – there is a teleradiology centre that receives emergency CT scan images from 50 hospitals simultaneously. Currently, there is no way that a radiologist can ascertain which CT scan is most relevant and hence, starts reporting them in chronological order – a scan done earlier is reported earlier. In comes an AI algorithm that can ‘read’ the CT scan even before the radiologist sees it in her/ his worklist. Eventually, a short while after the scan is acquired, when the CT scan does on the radiologist’s worklist, it is prioritised on the basis of the content of the actual scan. The algorithm may be able to detect, with a high degree of specificity, which scans are normal, and might de-prioritise them, making optimal use of the radiologist’s time and delivering a quicker result to the patient that needs it more.
When will AI take over the radiologists’ job?
This has to be the number one question on radiologists’ mind these days. Thoughts on this range all the way from ‘never’ to ‘tomorrow’. After having spent a considerable time studying the current state of AI both generally and as it relates to radiology, my views are a little guarded. Before imagining a future state where ‘AI takes a radiologist’s job’ it is important to know that most technological innovation comes in phases. There will not be a single day where AI algorithms take over radiologists’ jobs, there will be an incremental shift where radiologists will eventually become many times more effective. A radiologist who reports 20 MRI scans today will report 50 or 100 cases per day. A junior radiologist will be as good as a senior, more experienced one. And then, most importantly, there is the issue of liability, one which is curtailing the implementation of AI without human supervision. Who takes the blame if the algorithm goes wrong? Even the world’s most specific algorithm will one day label an abnormal case as normal, leading to some loss to a patient – who will take the blame for this? Will it be the software company that sold the algorithm? Will it be the organisation that bought it? Or the doctor who should have been there to monitor it?
(The author is unable to discuss specifics of most algorithms due to confidentiality and disclosure concerns.)