Use cases of AI in medical imaging
While currently quite niche, the areas of applications and use cases of AI in medical imaging are gradually becoming wide-ranging. Most of the work that AI is doing at the moment in medical imaging is around analysing. This means that when images are taken, AI algorithms are being developed that will read a particular image and look for any abnormalities before adding it to the radiologist’s worklist.
Some companies are doing exceptional work in other parts of the imaging workflow such as in the acquisition, assigning and reporting stages.
AI algorithms can look at the images and highlight any findings which deviate from the normal. This will help radiologists go through the images quickly, save their critical time and helps in clinical decision support. Moreover, AI has improved the way reporting is done. For example, a company called Nuance offers noteworthy capabilities in natural language processing for healthcare in general and for radiology in particular.
Source: Frost & Sullivan
So, when a radiologist starts dictating the findings, the company’s solution picks up all the dictation words and fills in the report by picking up the relevant words, so that the radiologist doesn’t have to spend time writing it. This becomes easy for subsequent algorithms to analyse a set of structured reports and then analyse it together to improve quality outcomes. Similarly, viewing and interpreting are other key areas where AI vendors are focused.
Today, AI is helping radiologists by guiding them towards factors that need immediate attention and increases their productivity. Therefore, it will not make their role obsolete but help them do more value-added work, which will increase their contribution towards clinical outcomes and operational efficiency.
Furthermore, AI can play a key role in low-resource settings. For example, in a developing country like India, tuberculosis cases are quite high and there may not be enough resources in the country or the ideal number of radiologists needed to go through all the X-rays required for tuberculosis evaluation. Or for instance, many countries have a mammogram screening programme, where women of a certain age group are asked to compulsorily undergo mammograms once every two or three years. This results in millions of different images. When you screen 100 patients probably three or four would turn out to be positive. In such a situation, the AI algorithm can look at those images first and detect the ones that deserve immediate attention.
In the past, whenever AI companies would approach hospitals, there were always questions about who is going to pay for this as there were no reimbursement models available. However, in the U.S., the Centers for Medicare & Medicaid Services (CMS) has approved reimbursements for two algorithms. One is for a diabetic retinopathy algorithm.
Many more such algorithms are likely to be approved for use in clinical settings. Also, as part of their business case, AI vendors can’t solely rely on reimbursement models but will need to show how much department time is being saved by the algorithms. This is when hospitals will join the AI adoption.
The combination of AI in radiology, pathology and genomics can improve diagnostic power but is at a very nascent stage.
The one area where it has got a bright future is oncology. The use case of the convergence of imaging and digital pathology and genomics to happen in oncology will lead to higher accuracy of diagnosis as well as earlier detection. Either of these two means that the health systems will save a lot of money.
Therefore, a hospital needs to identify the disease at the earliest stage and get the right diagnosis the first time. If this has to happen, hospitals have to rely on data from other disciplines.
Data from the other clinical disciplines is going to improve the clinical diagnosis and accuracy of imaging. This is currently termed radiogenomics, which is the combination of the fields of radiology and genomics.
The field of radiogenomics is currently overcoming the translational gaps that separate the pre-clinical and research environments where it is mostly established, such as in cancer research and drug development, from the clinical setting, where it is gradually gaining acceptance as a clinical decision-making tool.
When images are being fed to the AI algorithms, they will have some radiology findings and when these are correlated with the genomic or digital pathology findings, the AI algorithms can do the diagnosis with higher confidence and accuracy.
1. Digital Diagnostics (formerly IDx LLC): It is the first autonomous AI algorithm to be approved by the U.S. FDA. This means that the algorithm does not require a doctor to sign off on the diagnosis. The product is IDx-DR for diabetic retinopathy screening in diabetic patients.
2. Companies like RapidAI and Viz.AI have AI solutions that are being used in stroke diagnosis in hospitals, allowing for faster detection and communication to the clinical teams to ensure the fastest possible treatment, since every second in stroke cases counts.
3. When COVID-19 struck, several imaging AI solutions were developed and deployed to help doctors detect COVID-19 from lung imaging, to predicting the severity and prognosis of a positive patient (e.g., need for a ventilator), etc.
4. Lunit INSIGHT MMG is a commercial AI solution for breast cancer detection from mammograms. A study published in JAMA Oncology and Lancet Digital Health compared various commercially available mammography AI solutions with Lunit Insight MMG. Based on 8,805 cases, Lunit’s algorithm showed the best accuracy among other available solutions.
Nicholas Cnossen, MD, MPH and Manish Kohli, MD, MPH, MBA from Albright Stonebridge Group (ASG) highlighted that despite the potential of AI to transform healthcare, many pertinent issues are slowing more rapid and widespread adoption.
One of the most pressing issues is how to regulate AI in an accountable, fair, and transparent manner while implementing it for complex and risk-intensive processes.
AI regulation can be approached from an ethical, technical, or legal-regulatory perspective. There are various concerns regarding each of these approaches, and the concerns are often overlapping.
The ideal regulatory framework will likely require a multidisciplinary effort that adequately addresses concerns around data privacy, accuracy, patient safety and medico-legal implications. Within the legal-regulatory sphere, questions arise regarding the appropriate scope of legal policies surrounding AI as well as which entity or entities ought to be responsible for implementing and enforcing said policies.
There are accountability challenges because AI and deep learning algorithms such as neural networks are “black box” systems that defy traditional conceptualization and explanation. Individual errors or large-scale discriminatory patterns may result from AI for which it will be difficult to assign liability in the absence of a fundamental understanding of the underlying mechanisms resulting in error.