This website is intended for healthcare professionals only.
Take a look at a selection of our recent media coverage:
11th June 2024
Consultant radiologist Dr Gerald Lip took an early interest in the use of artificial intelligence in medicine. He tells Steve Titmarsh how his team’s work in Aberdeen is unveiling the technology’s potential to augment clinicians’ workflows in breast cancer screening, which they hope will improve patient outcomes and help colleagues manage an ever-increasing demand for their services.
It was a master’s degree in health informatics at Trinity College Dublin that ignited Dr Gerald Lip’s interest in artificial intelligence and its applications in medicine. And while his medical career began in surgery, after a short time he concluded that the world of radiology was better suited to him and so he commenced this training at Aberdeen Royal Infirmary.
Here, he was fortunate to be guided by an excellent academic team, including Professor Fiona Gilbert, now at Cambridge University, who led a number of national trials in breast screening and imaging, and was recently appointed the first clinical radiology AI lead advisor at the Royal College of Radiologists. It was this experience that brought Dr Lip to his speciality in breast screening, and after two years as a consultant he was appointed clinical director of the North East Scotland Breast Screening Programme.
In 2018, the agency Innovate UK wanted to establish five centres of excellence for AI, and one of these was based in Scotland, encompassing Aberdeen and Glasgow. Here, a five-year programme called iCAIRD – the Industrial Centre for Artificial Intelligence Research in Digital Diagnostics – was established and, for Dr Lip, the opportunity to determine the potential of AI in breast screening was a huge draw.
Dr Lip’s involvement in the iCAIRD radiology collaboration began with a retrospective study using a three-year dataset (1 April 2016 to 31 March 2019) from the Scottish Breast Screening Service.
Published in the journal Radiology: Artificial Intelligence in 2023, the results highlight the possible value of an AI tool to augment the work of human scan readers for breast screening.
This was only the beginning. Dr Lip and his team recently completed a prospective trial to evaluate the AI tool in a clinical setting, using it in addition to two human readers of breast scans in work funded by the NHS Health and Social Care Award.
The work was part of a project named GEMINI (Grampians Evaluation of Mia in an Innovative breast screening Initiative) – a collaboration led by the NHS Grampian Innovation Hub; Kheiron Medical Technologies, which developed the AI tool; and the University of Aberdeen.
The project uses AI to evaluate anonymised patient images in a secure and trusted environment for data processing. ‘We have a very strong information governance team in the hospital so what happens is screening images are stripped of identifiable patient information and given a pseudonymised number before being sent to the cloud. The AI does its analysis and then it comes back and within the NHS its rejoined [with the patient data],’ Dr Lip explains.
Some early initial results from GEMINI were presented at the European Congress of Radiology in March 2024. These revealed that by using AI as a ‘safety net’ image reader, 11 women were found to have cancer who would likely otherwise not have been detected until their next routine breast screening. As a result, they were detected at an earlier stage when their cancers were smaller.
This discovery has significant implications for breast screening, patient outcomes and health system resource. For example, ‘if a cancer smaller than 15mm in size is detected, generally the chance of survival is 95%’, Dr Lip says. ‘[A] key part of screening is we want to find small things before they become big things, and AI is one of the tools that can help us find these small things.’
In that way, this AI tool can act as a safety net, which can have positive implications for patients and reduce the need for more complex surgery, chemotherapy and radiotherapy.
Dr Lip’s team is currently working on a paper looking into the health economics of AI. ‘Similar work with chest X-rays by researchers in Glasgow shows that costs rise in the first two years of implementing AI due to greater detection rates,’ he explains. ‘But later, as the technology is scaled up, cost savings result from early detection and the reduced need for more expensive interventions for more advanced disease states.’
The next stage of this research will be to look at a larger dataset: Dr Lip’s team are co-applicants for an National Institute for Health and Care Research grant for a trial to evaluate AI in a UK-wide breast screening programme.
A key development in the evolution of breast cancer screening came in around 2015/16 in the switch from film to digital imaging. Traditional film imaging does not lend itself to the use of AI tools because the scanned film images are poor quality compared with digital counterparts. That development, coupled with the simultaneous rise of AI, led to their obvious pairing to enhance clinicians’ work.
The technology used in medical imaging, for example, is not the same as the much-discussed large language models such as ChatGPT. Instead, the AI used in the GEMINI project is very much a tool that can be fine-tuned to suit a specific use – and even a particular population – to produce the best results.
Another future benefit that Dr Lip hopes will result from employing AI is the automation of screening normal mammograms, reducing clinician workload. ‘Currently two radiologists look at each scan,’ he explains. ‘In future, if the AI records a mammogram as normal and a human screener agrees, there may be no need for a second human to verify the result.’
Indeed, initial results from an evaluation of AI using data modelling predicted a 30% reduction in clinician workload. Dr Lip explains that in a real-world situation this could mean his centre, where they read 20,000 mammograms a year, would not need a second human to read around 5,000 mammograms. This would free up a significant amount of clinician time for other tasks such as biopsies or face-to-face time with patients, he says.
Given that the Royal College of Radiologists estimates a 29% shortage of radiologists, which will rise to 40% in five years without action, and with breast screening likely to be one of the most affected specialties due to rising breast cancer cases in women, the potential time saving afforded by using an AI tool could be significant.
What’s more, as breast screening in Scotland sits under a national Picture Archiving Communications System (PACS), Dr Lip says scaling the GEMINI project to cover the whole country could be achieved ‘quite easily’. This could also be the case for Northern Ireland, which runs a unified PACS programme, as well as in England’s emerging imaging networks.
For Dr Lip, a key message from his research so far is that AI should be seen as supplementary and not a replacement for clinicians’ expertise – an experienced human still needs to be in the loop.
Of course, another reason for human involvement is that the AI only analyses the data it is given so it would not automatically know about previous surgeries and biopsies, unless this data was inputted. It may also unnecessarily recommend call-backs based on lesions from previous surgery, for example, whereas a human investigator would have access to the patient history and be able to exclude such scans from being re-read.
Dr Lip says that another significant factor highlighted was the importance of effective monitoring. AI systems suffer from drift, which happens when the software, machine or population changes, and it means that the performance of the AI responses slowly degrade over time. Monitoring is therefore needed to ensure the performance remains as expected and required.
And then there’s bias, particularly in terms of different patient populations. It is important to ensure there is no bias for different ethnic groups, for example.
‘Bias is a hot word in the AI world to ensure that it’s not biased against one group,’ says Dr Lip. ‘If an AI is only trained on white women, are there are differences in the mammograms of white women compared to black women? We don’t have significant knowledge to that level and it’s a fast-moving space.’
Real-world data from a representative population that the AI tool is analysing needs to be used to try to reduce biased results.
Looking ahead, Dr Lip says ‘more than one AI tool may be used to assess breast scans’, and he is aware of this currently happening in other clinical areas. ‘For example, in stroke care, there is a tool trained on the black areas of scans and a second trained on the white parts, so using the two tools in conjunction could improve results,’ he explains.
Another area of research for Dr Lip is in so-called masking. If there is a lot of glandular tissue in a breast, it can mask tumours and AI seems to perform well in these cases, he says. Colleagues at the Karolinska Institute in Sweden have developed a masking tool and a density tool, which can be used along with the AI screening tool, and there is more work to be done in this area.
Genetics is also key to revolutionising breast cancer care. At the Aberdeen Centre for Health Data Science, in collaboration with Professor Lesley Anderson and her team, Dr Lip is undertaking multidisciplinary work in this field.
Using BioBank data in breast cancer, the aim is to advance treatment personalisation. In future, by combining scan data with genetic and biopsy information, for example, it may be possible to determine how often each woman should attend screenings to ensure cancers are detected as early as possible.
With all their published work and a host of novel projects in the pipeline, it’s clear that the work Dr Lip and his team in Aberdeen are doing with AI promises to help alleviate workforce pressures in the NHS, detect more cancers at an earlier stage and, ultimately, improve outcomes for patients.
6th May 2022
Use of an AI-based breast cancer protocol has been found to have a similar screening sensitivity and a slightly higher specificity than radiologists and might therefore be able to considerably reduce the workload of radiologists.
This was the finding from a retrospective analysis by researchers from the Department of Computer Science and Public Health, University of Copenhagen, Copenhagen, Denmark.
Breast cancer arises in the epithelium of the ducts or lobules in the glandular tissue of the breast and according to the World Health Organization, in 2020 there were 2.3 million women diagnosed with breast cancer and 685 000 global deaths.
Population screening of women enables the detection of the early signs of breast cancer and one European analysis of observational studies concluded that the estimated breast cancer mortality reduction from invited screening was 25-31% and 38-48% for women actually screened.
Although screening mammography is the principle method for the detection of breast cancer, 10%-30% of breast cancers may be missed at mammography. Part of the reason for missing possible cancers may be due to behavioural factors.
For example, in one study, six radiologists who reviewed 100 breast cancer scans, where the prevalence of disease was artificially set at 50%, missed 30% of the cancers. In contrast, when the prevalence was raised, participants missed just 12% of the same cancers.
In other words, radiologists are more likely to be on the look-out of suspicious scans when they know that the disease has a much higher prevalence.
One potential way to remove the effect of behavioural influences is the use of an artificial intelligence (AI) based system for reading breast cancer scans. In fact, such systems have been shown to maintain non-inferior performance and reduced the workload of the second, radiologist reader by 88%.
But whether an AI-based breast cancer system could be safely used for population-based screening and reduce the number of mammograms that required reading by a radiologist is uncertain and was the objective of the current study.
Using a retrospective design, the Danish researchers examined whether the AI-based cancer protocol was able to detect normal, moderate-risk and suspicious mammograms. The team used data from a breast cancer screening program and each of the mammograms was scored from 0 to 10 (to designate the risk of malignancy) by the AI-based cancer tool.
The team then compared to AI-based cancer system and radiologists with respect to screening performance and used the area under the receiver operating characteristics curve (AUC) to compare performance.
AI-based cancer screening protocol performance
The cohort included 114,421 women with a mean age of 59 years who underwent mammography screening. The scanning identified 791 screen-detected cancers, 327 interval cancers and 1473 long-term cancers.
The AI-based cancer system had a screening sensitivity of 69.7% (95% CI 66.9 – 72.4%) which was non-inferior to the radiologist sensitivity of 70.8% (p = 0.02). The AI-based screening specificity was 98.6% and which was significantly higher (p < 0.001) than that of the radiologist (98.1%).
Based on these findings, the authors calculated that use of the AI-based cancer system led to a 62.6% radiologist workload reduction. Moreover, the AI-based system reduced the number of false-positive screenings by 25.1%.
They concluded that incorporation of an AI-based cancer system for population-based screening could both improve these programs and reduce radiologist workload and called for a prospective trial to determine the impact of AI-based screening.
Citation
Lauritzen AD et al. An Artificial Intelligence–based Mammography Screening Protocol for Breast Cancer: Outcome and Radiologist Workload Radiology 2022
IMPORTANT LINKS
MORE FROM COGORA
© Cogora 2024
Cogora Limited. 1 Giltspur Street, London EC1A 9DD
Registered in the United Kingdom. Reg. No. 2147432
