Artificial Intelligence (AI) is an innovative and potentially transformational tool in Radiology and diagnostics in general. The adoption and implementation of AI solutions within medicine has accelerated in recent years, with AI most commonly serving in a complementary role to the clinical workflow in order to expedite routine and well-defined processes.
These processes typically function as a clinical decision support system referred to as computer-aided detection (CADe) or computer-aided diagnosis (CADx). However, with the development of more complex deep-learning algorithms such as convolutional neural networks (CNN’s), AI has the potential to match or exceed human performance in domains involving vast amounts of information processing and synthesis. Given the profound and wide-ranging implications of AI implementation in healthcare, a foundational knowledge of AI is necessary for all pertinent provider domains. In this article, we will discuss the importance of AI systems and expound on current and impending utilisation of AI for Radiology and diagnostics.
Demand for AI in healthcare
The demand for radiologic diagnostics continues to grow and currently outpaces the growth of the Radiology workforce. For example, the Clinical Radiology UK Workforce Census 2020 Executive Summary estimated that the demand for complex imaging studies (CT, MRI) is increasing at a rate of 7 per cent per year while the Clinical Radiology consultant workforce is growing at a rate of 4 per cent per year. The same 2020 Executive Summary from the UK estimated that over 200,000 patients waited six weeks or more for a CT or MRI scan in September 2020, a number ten times greater than the year prior.
The COVID-19 pandemic has also created a backlog of patients requiring imaging studies. Part of this backlog is attributable to a sharp decline in utilisation of oncologic screening studies, with one study estimating a deficit of over nine million studies for prostate, breast, and colorectal cancer in the U.S. alone. The adoption and implementation of AI in Radiology has the potential to significantly alleviate this burden and could pave the way to more timely imaging and diagnosis for many patients.
In 2020, the ACR Data Science Institute Artificial Intelligence Survey showed that approximately 30 per cent of radiologists currently use AI in clinical practice, with more planning to purchase AI tools within the next five years. The survey found that large practices are more likely than small practices to use AI, and that the most common current usage of AI is for detection of intracranial haemorrhage, pulmonary emboli, or mammographic abnormalities. These AI applications are primarily clinical decision support systems (CADe, CADx) designed to supplement and expedite the workflow of a human diagnostician.
Established AI use cases
One of the more established use cases for AI in diagnostics is the in the detection of diabetic retinopathy on images of the retinal fundus. In one study, a CNN was developed and trained on over 12,000 images. This algorithm was able to achieve over 98 per cent specificity and approximately 90 per cent sensitivity as verified by practicing ophthalmologists. In another example, a deep learning algorithm was successfully developed and trained on over 14,000 three-dimensional optical coherence tomography (OCT) scans to make referral recommendations regarding range-of-sight threatening retinal diseases such as diabetic retinopathy which meet or exceed those of experts. In 2018, the FDA approved and granted Breakthrough Device designation to an AI system designed to detect diabetic retinopathy.
Breast cancer screening is an established but evolving use case for AI. CADe systems have been used for detecting mammographic abnormalities since as early as 1998 and have since become reimbursable through the Center for Medicare and Medicaid Services (CMS). Studies of CAD applied to mammography initially yielded mixed results with respect to accuracy and reading times. More recent studies with deep learning algorithms, however, have been much more promising. For example, in one study from University of Pennsylvania, a deep learning AI system was applied to digital breast tomosynthesis (DBT) breast screening studies with encouraging results including a 52.7 per cent decrease in reading time, an 8.0 per cent increase in sensitivity, and a 6.9 per cent increase in specificity as compared to radiologist performance without AI. Given these results and those from other similar studies involving deep learning, there is likely a substantially greater benefit in terms of accuracy and workflow optimisation from using more modern and sophisticated AI systems to assist with breast cancer screening.
Given that lung cancer is the most common cause of cancer-related death in the U.S., improvements in screening processes are crucial. For this reason, lung nodule screening with CT is another area in which AI has been implemented and shown positive results. One study used a deep learning algorithm to achieve exceptional CT screening performance with an area under the curve of 94.4 per cent. The deep learning algorithm performed at least as well as experienced radiologists in a direct comparison.
Transformative AI use cases
Recently, AI has been applied to the diagnosis and prognostication of COVID-19. Use cases involving COVID-19 have been particularly important due to overwhelming strains on medical infrastructure caused by an unprecedented volume of patients. In one study out of Stanford, a CNN was developed to detect COVID-19 on chest CT without any special image processing. This CNN is expected to additionally enable prognostication by tracking image features over time on sequential scans.
In another example, a deep learning model was developed to output biomarkers derived from CT scans from institutions around the world. These derived biomarkers were then utilised with electronic health record (EHR) data for prognosis analysis to identify predictors for patient severity outcomes.
Barriers to AI implementation
Despite the potential of AI to transform healthcare, there are many pertinent issues which are slowing or preventing 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. As explained by Corinne Cath, 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 address 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 exist significant accountability challenges due to the notion that AI and especially deep learning algorithms such as neural networks are “black box” systems which defy traditional conceptualisation and explanation. Individual errors or more large-scale discriminatory patterns may result from use of AI for which it will be difficult to assign liability in the absence of a fundamental understanding of the underlying mechanisms resulting in error.
Currently, reimbursement for AI by the U.S. Center for Medicare and Medicaid Services (CMS) is limited to use in mammography. Additional uses of AI software deemed reimbursable by CMS will likely depend on evidence showing significant improvements in diagnostic accuracy and operational efficiency. For now, healthcare systems must assess the potential for AI to impact operational efficiency in order to evaluate potential cost-savings.
According to a recent analysis by Gartner, the AI investment landscape appears positive. This analysis indicates that, in 2020, 47 per cent of AI investments were unchanged and 30 per cent of organisations planning to increase their AI investments despite the impact of the COVID-19 pandemic.
The hype cycle for AI in 2020 showed a few new AI trends debuting on the Gartner Hype Cycle, including generative AI, composite AI, and responsible AI. Most AI trends reside near the Peak of Inflated Expectations, with AI as a general concept felt to be rolling off the Peak of Inflated Expectations as it begins to deliver on its potential by providing benefits for businesses.
Future of AI
The perceived lack of trustworthiness in current AI systems is related to the nature of complex AI algorithms as potentially unexplainable, a notion which confers significant reluctance to accepting and utilising the output generated by AI. To address this concern, developments are being made in the field of Explainable AI (XAI). One XAI study performed on the Heart Disease Dataset from UCI showed multiple approaches to generating XAI, including feature-based techniques showing how much input features contribute to model output as well as example-based techniques showing how much individual instances contribute to a model’s output.
Over time, AI algorithms, not only in radiology but healthcare generally, are bound to establish a key niche and find broader acceptance helping impact workflow efficiencies, safety, quality access and cost partly through automation of low-level tasks, better data organisation and contextualising insights. In order for society to realise this technology’s full potential commensurate regulatory, medico-legal and payment reform will be key.
Reference available on request.
This article appears in the latest issue of Omnia Health Magazine. Read the full issue online today.