AI and Humans: Working Together to Improve Diagnosis

Earlier this week, UK Prime Minister Theresa May announced plans to leverage the valuable role that artificial intelligence (AI) can play in transforming the diagnosis of cancer and other diseases. One of the goals of the government's industrial strategy is the development of algorithms which could be used to detect cancer earlier, preventing avoidable deaths occurring from late diagnosis.  

“The NHS has practically unused archives of millions of diagnostic images that could become one of the most powerful clinical datasets in the world if artificial intelligence is used effectively," notes Jane Rendall, managing director of NHS imaging technology partner Sectra UK & Ireland.

To learn more about the impact AI has already had on diagnostics and health services, as well as some of its future capabilities, we spoke to Chris Scarisbrick, former radiographer and now Sales Director of Sectra.

Anna MacDonald (AM): Which areas of diagnostics is AI already established in?

Chris Scarisbrick (CS): AI is already being put to good use across a range of Diagnostic Disciplines, the last time I looked there were well over 100 start-ups operating across a range of specialities ranging from lifestyle management & monitoring, mental health, insights and risk management, etc. Medical Imaging however, is particularly well suited to the application of AI. The wealth of data available in terms of medical images, combined with the fact that the AI ‘learns’ in much the same way as a Radiologist/HistoPathologist -through pattern recognition - means that AI is particularly well suited in assisting with the diagnosis of conditions normally diagnosed with the human interpretation of imaging data. 

Major players such as Google, IBM and Intel are all investing heavily in Healthcare based AI development, not to mention the plethora of small start-ups, all of which with their own roles to play.   Imaging techniques are becoming more and more sophisticated, invariably producing exponentially more data to interpret, I have heard it compared to “Trying to drink water from a fire hose”. This, at a time when our highly qualified and skilled experts are becoming increasingly scarce. For these reasons, most AI development is aimed at making our valuable clinical resources more efficient.  Some examples of AI in healthcare diagnostics today (by no means exhaustive):

a. Early detection of Osteoporosis

b. Automated tagging of lesions in CT scans

c. Breast imaging reads – suggested lesions on a first read

d. Assessment of liver scans 

e. Identification of glaucoma from retinopathy imaging

f. Automated cell counting and ki67 assessment of histopathology slides

g. Etc

h. etc

AM: What impact has AI technology had on health services such as the NHS, and patients?

CS: Already, in certain disciplines we are seeing report turnaround times reduced.  Radiologists and Pathologists are benefiting from tools which reduce their time to report complex imaging by half (and sometimes more) by automating some of the processes that would previously have been manual. This means that patients get their results in a more timely fashion and helps to limit bottlenecks for our NHS organisations. This is especially important for the patient whom is awaiting a crucial cancer result on a liver biopsy (for example).  In these time critical conditions any time saved in the diagnosis improves the outcome for the patient. 

Additionally, AI is frequently being used to flag early warnings to patients. Medical imaging, family history, lifestyle factors, genetic data etc is now all being combined to assess risk factors for individual patients for certain diseases. This means that the NHS can focus more on the prevention of these conditions in the future by medicating or making lifestyle changes today.  Moving healthcare models into preventative rather than reactive modes could save the NHS significant cost as the treatment of chronic diseases is expensive.

The NHS five-year plan identifies a funding gap, efficiency gains are crucial if this gap is to be closed. AI is seen as a key enabler to allow the NHS to do far more with less.

AM: What challenges need to be overcome before more widespread adoption of AI in diagnostics?

CS: Access to high quality datasets is one of the key bottlenecks in the development of AI applications. The algorithms ‘learn’ through the analysis of huge quantities of data and this data needs to be ‘noiseless’, structured, tagged, and anonymised. The vast majority of the data within the NHS does not meet this criteria so work needs to be done to make high quality data sets available to the developers of these applications. The sub-specialists that are needed to perform this data cleansing are already at capacity dealing with the overwhelming amount of clinical reporting that they need to get through in their day job.

AM: How reliable is AI, and how accepting are healthcare professionals/patients of its abilities?

CS: In some areas, AI can already outperform the human.  Especially when dealing with very common conditions where large amounts of ground truth data is available. This is where we will see AI becoming more and more important, certain conditions can be triaged and prioritised by the computer to allow our human resources to be used on the less common conditions where the AI is far less effective. However, in almost all example use cases, a ‘human in the loop’ philosophy produces the very best results.  Utilising AI to make suggestions, highlight suspicious findings and automate some of the analysis often produces much higher specificity than either the AI alone, or the human alone. A phrase has been coined within the industry “A Radiologist that uses AI will be better than a Radiologist that does not use AI”. The research certainly supports this standpoint.

There seems to have been a major paradigm shift these days. AI is everywhere, flying our planes, driving our cars, and even choosing what we watch on the TV. This means that patients are much more accepting (even expecting) AI to play a part in the healthcare moving forward. Many healthcare professionals see it a necessity these days to getting their job done, especially when the academic studies provide clear, peer reviewed evidence for the success of these algorithms.

AM: Are there any areas of diagnostics that AI is not suited for? Will it ever completely replace humans?

CS: Diagnosis is always based on the interpretation of data in one form or another. For this reason, the application of AI will almost always have the potential to improve any area that it touches. However, all studies suggest that the best results are obtained when the human is in the loop. Augmenting the workflow of our diagnosticians using AI applications will drastically improve efficiencies and accuracy of the clinical processes. Thus, improving patient outcomes and helping to close the reporting gap that currently exists due to the lack of trained professionals combined with the ever-inflating mountain of data for them to interpret. It is unlikely to completely replace humans, but it will allow humans to be more effective. 

Chris Scarisbrick was speaking to Anna MacDonald, Science Writer for Technology Networks.