Artificial Intelligence and Mammography
Dr Justus Apffelstaedt, specialist surgeon with an interest in breast, thyroid and parathyroid health as well as soft tissue surgical oncology.
There has been an almost deafening buzz in the media about artificial intelligence (AI) in recent months. This buzz has been driven by the application of AI in popular search engines such as Google and the integration of it into everyday software programs such as Microsoft’s Office suite. Spectacular images generated with the help of AI have furthered public fantasies.
However, much less attention has focused on AI applications in medicine.
Image interpretation is one of the areas where AI has been at the forefront of computerisation, with particularly mammography (essentially a two-dimensional image with only shades between black and white) a prime target for computerised analysis.
I saw the first computerised analysis of mammography in 1996 at a visit to a start-up in Silicon Valley. At the time, images were still film-screen and first had to be digitized and then read into a computer before being analysed. This was a clumsy, time consuming and labour-intensive process; accuracy was limited. But it was immediately clear to me that, with further development, this was the future.
Fast forward to 2023 and several important trends have changed the landscape of medical imaging. Cross-sectional imaging has become the norm for many of our body’s organs and, with it, a flood of images to be analysed. CT scans, MRIs and PET scans typically generate (per examination) more than 100 images to be analysed. Mammography, with the advent of tomography, also now consists of a series of more than 20 images.
This has led to an exponentially rising demand for imaging readers world-wide. At the same time, resources are constrained; the more so in a developing country with its well documented and increasing shortage of specialists across all fields. In parallel, computing power and with it image analysis software has improved by leaps and bounds.
Now, for the first time in a major trial, AI enhanced mammography interpretation has been shown to be at least as good as a human reading. Published in the pre-eminent journal ‘The Lancet Oncology’ this August, Kristina Lang and collaborators from Sweden report on the results of AI supported versus standard human double-reading of screening mammograms in 80000 women.
This is a major break-through for women. We have run computer assisted mammographic screening for more than a decade, but previous systems were far from perfect. In screening mammography, there are two major problems. One is, obviously, the missed breast cancer. The second, much less present in the public mind, is the false positive (charmingly termed ‘erring on the side of caution’). This is where a patient is told to come back for reassessment in 6 months or where a biopsy is performed, and the lesion is found to be benign.
In about 90% of the second opinions, we see for assessment of mammographic lesions, we could reassure these extremely anxious and distressed women that no intervention is needed. For the last couple of months, we have been working with latest generation AI enhanced mammography interpretation. The difference is remarkable in that prior systems had way too many false positives while, on the other hand, missed some subtle cancers. The Swedish study shows that there was actually a trend towards more cancers being detected in the AI enhanced group versus the human readers with no difference in the number of false positives.
This has also been our experience in practice.
There are two other advantages of AI enhanced mammography interpretation that are less obvious but of utmost importance. In screening mammography, reader fatigue is a real problem. In 1000 screening mammograms there will be between 5 and 10 cancers hidden. This means that 990 to 995 mammograms are normal or show benign lesions. Computerised systems don’t know fatigue, certainly have no ‘post-lunch slump’, don’t get sick and work much faster than human readers.
We have not seen any time delay between the image taken and available for viewing. Another advantage is that it takes more than 10 years to train a specialist doctor before (s)he can interpret mammography at the required level for organised screening programs. AI based systems can be trained within weeks or even hours.
Currently these systems are very expensive and as such will be the preserve of specialised centres with the necessary volume to justify the expense. More than a quarter of a century after my first exposure to computer assisted mammography interpretation, AI enhanced mammography interpretation has arrived in practice. This is the advance that should be offered to all women who go for mammography.
Finally, a glimpse into the future. We predict that with the new systems a human reader is still required to make the final decision but, in due course, this requirement will probably become obsolete.
Biography – Dr Justus Apffelstaedt
Dr Justus Apffelstaedt is a former Professor of Surgery and Head: Surgical Oncology Service, University of Stellenbosch. Dr Justus Apffelstaedt earned a Medical Degree and a Doctorate in Medicine in Germany, as well as an MMed and FCS(SA) in South Africa and an MBA from Bond University in Australia.
He has represented developing countries on the Breast Surgery International (BSI) council and is a founding member and first chairman of the Breast Interest Group of Southern Africa (BIGOSA).
He is a fellow and life member of the International Union Against Cancer (UICC) Fellows.
He is excellent at translating complex medical terminology into easily understood language and is a proponent of proactive breast health management through extensive dissemination of information to the general public.
His breast service is the only one in Africa to publish peer-reviewed data comparable to international breast practices in breast screening. He is also the author and co-author of several publications in peer-reviewed national and international journals on breast cancer screening and breast health issues.
His current interest and field of practice includes breast health. thyroid, parathyroid and soft tissue tumours.
He has been proudly supporting one of South Africa’s oldest, national non-profit breast cancer support groups, Reach for Recovery, to raise awareness and funds for their Ditto Project since 2018.
He has practices in Cape Town, South Africa and Windhoek in Namibia.