How AI software is helping spot breast cancer, especially early stage cancer, to save more lives by hastening treatment – doctors explain its value, and potential
- Artificial intelligence can find patterns in mammograms that humans are not able to spot and thereby catch early stage breast cancer, Hong Kong radiologist says
- For all the promise of AI, doctors remain cautious and say human expertise is still needed to ensure the most accurate diagnoses and successful outcomes
Every day in Hong Kong, an average of 15 women are diagnosed with breast cancer – the most common cancer affecting women in the city, and the third leading cause of cancer deaths among them. It accounted for 28.5 per cent of all new cancer cases among women in the city in 2021.
Worldwide, breast cancer is the most frequently diagnosed cancer and the leading cause of cancer death in females: an estimated 2.3 million women were diagnosed with breast cancer in 2020 and around 685,000 died from the disease that same year.
Discovered in its early stages, however, breast cancer is highly treatable.
The key to early detection of breast cancer since the 1980s has been mammography screening. Differences in radiologists’ diagnostic accuracy often leads to unnecessary return visits, or worse, missed cancer. There is also a shortage of breast radiologists worldwide. Artificial intelligence (AI) has the potential to overcome these challenges.
Numerous studies have shown that AI may perform as well as, or better than, humans in tasks such as detecting malignant tumours in medical imaging, and triaging cases – the process of assigning priority to patients according to the urgency of their need for treatment.
In Hong Kong, the Hospital Authority has already implemented AI to help triage and detect abnormal findings in chest X-rays.
Why regular mammograms can be lifesavers: they spot breast cancer early
Now Hong Kong radiologists – following in the footsteps of peers in Sweden, Hungary and United States, among other places – are beginning to test and use deep-learning AI software tailored for breast imaging, given its potential to increase detection rates for breast cancer and make the process of interpreting mammograms more efficient.
“Artificial intelligence is very exciting, especially for us in the radiology field,” says Dr Alta Lai Yee-tak, a consultant radiologist at Pamela Youde Nethersole Eastern Hospital in Chai Wan and honorary secretary of the Hong Kong College of Radiologists.
She adds that although computer-aided detection (CAD) has been used in mammography for over 20 years, what makes AI different is its potential to learn.
“There are huge data sets that we can potentially feed into AI, and it can potentially gain a lot of insights,” she says.
“With deep learning, AI may potentially even find patterns that we as human beings are not able to spot right now – it might be able to analyse mammograms in a way that we as human beings do not know how to. So there is a lot of promise in that.”
Radiologists Dr Julian Fong Chun-yan and Dr Lui Chun-ying, who co-founded the breast screening centre Hong Kong Women’s Imaging (HKWI), began testing the Lunit Insight MMG, an AI solution for mammography, in 2022.
In January 2023, they began integrating the software, which has been approved by the United States’ Federal Drug Administration, into their centre’s workflow, and since then, HKWI has processed more than 10,000 mammograms using the software.
The centre is the first and only breast screening centre in Hong Kong to formally use AI on a daily basis – the Hospital Authority has not yet approved AI mammography solutions for formal use, although radiologists at public hospitals have begun to use AI software on a trial basis.
There are other solutions in the market, including ProFound Breast Health Suite’s ProFound Detection, from US-based iCAD; UK-based Kheiron Medical Technologies’ Mia Reader; and Netherlands-based ScreenPoint Medical’s Transpara.
Fong says Insight MMG is especially suitable for use in Hong Kong because 145,000 of the 170,000 sample cases that the software was developed with came from South Korean women (Lunit is an AI-based medical imaging equipment firm based in South Korea).
“Their machine learning samples are mostly Asian. That’s very important for us, because breast composition – it’s well known – is quite different between Asians and Caucasians,” Fong says.
Asian women tend to have a higher proportion of dense breast tissue, which can make cancers more difficult to detect in a mammogram, says Dr Hung Wai-ka, a surgeon who specialises in breast surgery and is a member of the Hong Kong Breast Cancer Foundation’s management committee.
The most widely used screening tool for breast cancer is the 2D mammogram; supplementary tools and imaging – such as ultrasound and 3D mammograms – are often used for women with dense breasts to improve cancer detection.
“The way I see AI in the screening of breast cancer is very much like a supplementary tool,” Hung says.
According to Hung, studies have shown that when AI technology is used, more early stage cancer can be detected.
He says: “When we detect the cancer at its early stage, the treatment is much less invasive – less surgery, less surgery [on] the lymph nodes, and also less chemotherapy. Also the overall treatment cost will be much less.”
If AI proves to be continuously reliable, it can not only potentially help doctors identify cancer with increased accuracy and at earlier stages, but also allow patients deemed a lower risk to avoid needless anxiety and unnecessary biopsies, Fong says.
“After a few similar cases in my patients, I already trust it a lot,” he says of the AI software.
Despite the potential, Fong cautions that the AI software is not perfect, and that experienced breast radiologists are still needed to interpret findings correctly.
In one case, the AI software had indicated that a particular calcification was of concern. But upon examining the mammogram in another view, Fong determined that the calcification was over the skin, and therefore could not be cancerous.
While calcification can be an indicator of cancer, most is of benign origin – fibroadenosis (benign breast lump), hormonal stimulation, and even lotion on skin can cause calcification.
“Also, for radiologists, we will review all the images, not only the single images,” Lui says. Because radiologists have access to previous imaging, they can see if an area of concern has been stable for a few years and is therefore benign, whereas AI in its present state would not be able to make that judgment.
Hong Kong start-up’s faster, cheaper MRIs new option for breast scan
Doctors also review patients’ clinical information, age, race and gender, and individual risk when interpreting results.
“AI cannot replace humans at this juncture, because we need to have other clinical background, our experience, to assist us to make the diagnosis,” Fong says. “We have to work together with the computer. Human and computer, we are complementary.”
Hung agrees. “[AI] is just alerting the radiologist [to] the area of concern, and it is the radiologist who makes the final decision regarding whether [the] particular area highlighted is cancerous or not, whether further action should be taken.”
Lai urges caution when analysing research about AI mammography solutions, pointing to a study published in medical journal The Lancet in August 2023 that involved over 80,000 breast screenings in Sweden.
‘Get the damn mammogram’: breast cancer awareness month
“They randomly separated the [women’s screenings] into two groups. One of them went through AI software first. If AI deemed them to be low risk, or to have low likelihood for having cancer, then those mammograms got read by just one radiologist,” Lai says.
“Then the other group just went through the usual standard of care, which is having two radiologists read every single thing. And the results are very exciting. They say that cancer detection rates between the two groups are more or less similar.”
Because the study was published in August 2023, not enough time has passed to do a two-year follow-up to confirm that the AI diagnoses were correct, Lai says.
“We need that two-year follow-up to be sure that the women they say are negative are truly negative. We want to wait two years – that’s what we do for breast cancer trials,” she says.
Training the AI is also important, to ensure that the information it learns from reflects real life.
“There is also a possibility of amplification of biases, for example misinformation from the training data set,” she says. “If we teach that to a computer or to AI, those kinds of biases or mistakes might get amplified.”
Despite the associated challenges, Lai is optimistic about AI’s future in the field, as are the other doctors.
“For breast screening, we hope to pick up early cancers … to have early diagnosis and treatment,” Lui says.
I asked experts how best to prevent breast cancer after my own mass was found
“Whenever there is any new technology that can help with this, it is our hope that we could adopt it to help more patients. … [AI is] not perfect at the moment, [but] the technology will improve more and more.”
Lai adds: “Together with AI, human beings can perhaps do a better job in terms of healthcare outcomes.”