AI in Radiology: Aiding Workflows and Accuracy as Workforce Pressures Mount-The HSB Blog 9/30/23

Our Take:

Artificial intelligence has numerous applications in radiology and has been rapidly evolving to help improve care, reduce costs, and reduce the burden on radiologists. AI algorithms are being developed to assist radiologists in the analysis and interpretation of medical images and can help identify abnormalities, quantify tumor sizes, and highlight potentially relevant areas for further review. AI also can help automate time-consuming tasks in radiology, such as image segmentation and feature extraction which can significantly reduce the workload for radiologists, allowing them to focus more on complex cases and patient care.

This is particularly important as there is already a worldwide shortage of radiologists, which is projected to worsen as the population ages. For instance, in the U.S., the growth of the Medicare population has significantly outpaced the number of radiologists entering the field in recent years. As noted by one study presented at the Radiological Society of North America (RSNA), “the growth of the Medicare population outpaced the diagnostic radiology (DR) workforce by about 5% from 2012 to 2019” and there are no signs of this imbalance improving given that “between 2010 and 2020, the number of DR trainees entering the workforce increased just 2.5% compared to a 34% increase in the number of adults over 65.” As such, AI has the potential to revolutionize radiology by improving the speed and accuracy of image analysis and improving the quality of care. However, its adoption should be carefully managed to ensure patient safety and the continued involvement of radiologists in the decision-making process.

Key Takeaways:

Between 2010 and 2020, the number of diagnostic radiology trainees entering the workforce increased by just 2.5% compared to a 34% increase in the number of adults over 65 (RSNA)
In one study, comparing AI-CAD and traditional CAD software, the AI system outperformed by decreasing the false-positive marks per image (FPPI) by a significant 69% (Diagnostics)
Over 85% of outpatient facilities and hospitals are facing staffing challenges, while they’re anticipating a 10% uptick in demand for staffing across MRI, nuclear medicine, ultrasound, radiologic and cardiovascular technologists (U.S. DOL)
In one study from the Netherlands of over 40K women with extremely dense breast tissue scanning using commercially available AI software led to significantly fewer interval cancers than the control group (Pediatric Radiology)

The Problem:

Radiology is a true early adopter of AI in clinical practice. There were 520 FDA-cleared AI algorithms that were cleared as of January 2023 over three-quarters of which were for radiology. Nevertheless, several challenges need to be addressed to further broaden AI’s integration into healthcare even deeper. For example, while AI could eventually eliminate the need for additional readings or verifications by other radiologists, these algorithms need to be validated and tested in clinical practice before organizations will actually put them into practice and trust them.

One issue is the dual problems of data quantity and data quality. AI algorithms require large volumes of high-quality data for training and validation. However, as pointed out in a summary of an RSNA-MICCAI Pane entitled “Leveraging the Full Potential of AI—Radiologists and Data Scientists Working Together” older images may have certain idiosyncrasies. For example, while, "there was no good reason for it, a small percentage of the cases had text burned into the images, information such as dates, and computed radiography cassette numbers and other researchers…[were] still running into issues because they’re going back to older data that may have burned-in text." Also, as with any AI data set, developers of algorithms need to pay attention to bias in data training sets and in model output. Those building AI algorithms need to ensure they obtain access to diverse and representative datasets, which can be challenging, as data may be fragmented across different healthcare systems and data privacy and security must be ensured.

In addition, as noted in “Legal consideration for artificial intelligence in radiology and cardiology”, ”there is not a good regulatory framework for AI in the U.S. …there is no guidance on how to deploy the technology safely and there are no clear protections from lawsuits”. While the FDA released the “Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD) Action Plan” in 2021, this acts more as a framework and is based on total product lifecycle, it does not provide guidance on who should be responsible for product or software malfunctions that could impact patient care, resulting in misdiagnosis or even death.

Moreover, as highlighted in the “Legal considerations for AI…” article, the technology team needs to understand AI’s clinical impact and exactly where and how the AI can impact liability. For example, it is important to map out “all the places where error can occur” and all the “points of potential failure” including installation, server maintenance and procedures to ensure the AI is running correctly. In addition, the article stresses the need for vendors and enterprises implementing AI solutions to have validated methods to monitor and test the algorithms. Moreover, given the inconsistency of state laws and jury interpretations, Brent Savoie, M.D., J.D. Section Chief of Cardiothoracic Imaging for Vanderbilt University Medical Center recommends, “if there are not specific regulatory protections for an AI vendor or healthcare groups using the AI, they may think twice about implementing the AI or doing business in that location.”

Another issue will be ensuring the pace of regulation keeps up with the pace of technological change. Clear guidelines and standards for AI in healthcare need to be established to ensure patient safety while promoting innovation. As pointed out in “What's next for AI regulations in medical imaging” by the imaging platform Interlad, “one of the main challenges is that the FDA's traditional regulatory framework and review processes are not designed to keep pace with this speed of innovation, as AI-enabled medical applications are evolving rapidly, sometimes in unanticipated ways.” As such, models then need to be validated for accuracy, reproducibility, and applicability to the clinical problem they are trying to solve often in near-real time, all within the context of relevant data privacy and security laws.

The Backdrop:

Radiology generates vast amounts of data, and using AI has the potential to reduce read times, improve accuracy, decrease workforce burdens and even pinpoint new or earlier treatments. For example, in an article entitled “How does artificial intelligence in radiology improve efficiency and health outcomes?” the authors point out that, “AI could contribute to this in clinical, but also non-clinical, ways…even before a patient enters the radiology department, AI software might aid the scheduling of imaging appointments and predict no-shows for nudging or more efficient scheduling”. In addition, the article points out that “The workflow might also be optimized by changing the diagnostic process with AI, [for example in mammography screenings] studies have been performed to simulate an alternative workflow in which an AI risk score determines the number of radiology reads (none, single or double), reducing the total amount of reading time.”

In addition to reducing the sheer number of images to review, AI and other improvements like computer-aided detection can help decrease the time radiologists spend reading scans. However as the authors highlight, ”besides the quality of the AI system, workflow integration is crucial for making this kind of software a success.” but once that is achieved there can be dramatic improvements in efficiency. For example, “the automated quantification of nodules, brain volumes or other tissues…might mitigate some of the tedious manual work that is part of a radiologist’s job, along with the large interrater variability inherent to these tasks.” Similarly, when AI is combined with computer-aided design systems, AI workflows can be dramatically streamlined. For example, as the authors highlight in “Redefining Radiology: A Review of Artificial Intelligence Integration in Medical Imaging”, “AI-CAD systems’ merit lies in their substantial reduction of false positives, enhancing dependability in clinical settings. [In one study] comparing AI-CAD and traditional CAD software…the AI system outperformed by decreasing the false-positive marks per image (FPPI) by a significant 69%.[The article added] it specifically excelled in identifying microcalcifications and masses, reducing false positives by 83% and 56% respectively”. Clearly, properly trained and implemented AI radiological systems can lead to meaningful improvements in diagnostic accuracy.

AI radiology can combine these advancements in early detection to help develop precision treatments. For example, as demonstrated in “How does artificial intelligence in radiology improve efficiency and health outcomes?” by analyzing subtle patterns in medical images that might be missed by human observers. Utilizing the case of women with dense breast tissues, the article noted that screening can be personalized for this group, which is known to have a higher risk of breast cancer. The authors note that "a study from the Netherlands involving more than 40,000 women with extremely dense breast tissue [were scanned using commercially available AI software] resulting in significantly fewer interval cancers” than the control group.

These factors collectively create a fertile ground for the development and adoption of AI in radiology, with the potential to significantly improve patient care, enhance diagnostic accuracy, and optimize healthcare delivery. However, it's important to address the challenges and ethical considerations associated with AI in radiology to ensure responsible and safe implementation.

Implications:

As noted above, AI has significant potential to help increase efficiency and improve workflows all the way through the radiology process. As noted in one study, AI can have an impact “beginning from the time of order entry, scan acquisition [all the way through to] applications supporting image interpretation… and result communication” This is extremely important as the volumes of imaging scans have increased dramatically in recent years, with research indicating that due to “an aging population and a greater reliance on imaging in the United States and Canada, there has been significantly increased computed tomography (230%), magnetic resonance imaging (304%), and ultrasound (164%) imaging use within the last 2 decades” Importantly, this is occurring against the backdrop of fewer radiologists being available to read scans. As Radiology Business stated in an article entitled “10 trends to watch in diagnostic imaging” according to the U.S. Department of Labor Statistics “over 85% of outpatient facilities and hospitals are facing staffing challenges, while they’re anticipating a 10% uptick in demand for staffing across MRI, nuclear medicine, ultrasound, radiologic and cardiovascular technologists.”

In addition, while it will be crucial to continue to monitor, validate and audit any AI-based or assisted radiology system for bias, ethical issues and data security, AI algorithms can learn and adapt over time, potentially improving their performance and application for prevention, diagnosis and treatment of disease. Perhaps most importantly, AI in radiology can empower timely and accurate diagnoses which when coupled with personalized treatment plans, can lead to improved patient outcomes and quality of life. This is especially significant in conditions with high morbidity and mortality rates such as cancer.

Related Reading: