LONDON – The U.K. National Institute for Health and Care Excellence (NICE) has published new advice on how and when artificial intelligence (AI) could be applied to the interpretation of mammograms and chest computer tomography images, in a move that is intended to set the ground rules for the uptake of these technologies.
In population breast screening, NICE looked at how five AI systems could be used to pick out mammography images that need further assessment, supporting qualified radiologists in their interpretation.
Assessors commissioned by NICE reviewed evidence from 71,470 mammograms, concluding, “AI technologies may improve performance and save time in interpreting mammography exams.”
However, there are uncertainties, because the AI systems looked at in the review had been trained on images from women who did have breast cancer. That is not representative of the target population in the national breast cancer screening program, which includes all women aged 50-71, who are invited for screening every three years. A total of 2.4 million women were screened in 2017/18.
Currently, all mammograms are read by two radiologists, and referred to a third if they disagree. A government commissioned review of adult screening programs published at the end of 2019, suggested that “within a few years” AI will have been sufficiently well tested on the target population to replace one of the human readers. The Royal College of Radiologists has endorsed that view.
NICE agrees workloads could be reduced by replacing one of the readers, and also suggests AI could be used to triage images according to the likelihood of malignancy, so that these mammograms can jump the queue and be referred to a human expert sooner.
In addition to speeding up interpretation, NICE says the evidence it considered in drawing up its guidance suggests AI can pick up changes, such as microcalcifications and asymmetry density that are small and difficult to interpret by eye, even for experienced readers. Use of AI could reduce the number of false positive and false negatives.
Advice for local commissioners
NICE’s summaries of the use of AI in interpreting mammograms and chest CT images are published in Medtech Innovation Briefings. Unlike other opinions from the health technology assessment body, these are not instructions to local health commissioners that a technology should be adopted, but rather advice notes to support commissioners in decision making. The aim is to provide fast summaries about the capabilities of innovative devices and technologies, to avoid local commissioning groups having to do their own appraisals.
It is also the case that NICE does not have responsibility for deciding what technologies will be used in national screening programs. That is the role of the U.K. National Screening Committee (NSC), which is required to review any changes to how screening is conducted.
In March 2019 NSC issued guidelines on what evidence it will require to approve adoption of AI in breast screening, in which it too highlighted the shortcomings identified by NICE, in that algorithms have not been trained on relevant data, there are no prospective studies, and no head-to-head comparisons of different vendors’ systems.
Despite the holes in the evidence base, there is broad support for the use of AI, including giving companies access to data on which to train their algorithms.
A large government-funded study run by the East Midlands Radiology Consortium, whose members include 11 eleven hospitals, has just completed a two year long assessment of the use of AI and machine learning to improve population level breast cancer screening. Its final report, published in December 2020, made a number of recommendations on further research, and the training and IT infrastructure required to support adoption.
The situation is somewhat different in AI for interpreting CT chest scans, in that the systems NICE considered are all CE-marked. Rather than being assessed in the context of population screening, NICE assessors looked at their use in diagnosis of lung cancer, venous thromboembolism disease, respiratory conditions, chronic obstructive pulmonary disease, tuberculosis and idiopathic pulmonary fibrosis.
Despite the length of this list, NICE says currently there is limited evidence to support the use of AI in chest CT and more studies of the impact on clinical management are needed.
COVID-19 pushes the AI accelerator
As in so many other areas, from the development of vaccines to the deployment of digital health, the COVID-19 pandemic has prompted a shift in gear in the use AI as a tool in health care.
NHS has just made available 40,000 CT scans, MRIs and X-rays from more than 10,000 COVID-19 patients, for use by hospitals and universities across the country to develop AI solutions to tackle the virus.
It is hoped the National COVID-19 Chest Imaging Database (NCCID), will speed up diagnosis, ultimately leading to quicker treatment. The database also could help inform the development of a national AI imaging platform to collect and share data required to develop AI technologies for use in other indications, including cardiovascular disease and cancer.
Clinicians at Addenbrooke’s Hospital in Cambridge are using NCCID to develop an algorithm to inform more accurate prognoses of patients when they present with potential COVID-19. Using visual signatures of the virus in chest scans, they can compare patterns in an individual patient image with those in scans held in NCCID, to get a better idea of how a patient will progress.
“NCCID has been invaluable in accelerating our research and provided us with a diverse, well-curated, dataset of U.K. patients to use in our algorithm development,” said Carola-Bibiane Schönlieb, head of the Cambridge Image Analysis group at Cambridge University. “By understanding in the early stages of disease whether a patient is likely to deteriorate, we can intervene earlier to change the course of their disease,” she said.