AI – an innovative solution to breast cancer detection

Radiologist Dr Axel Gräwingholt, Clinical Co-Chair of the Guidelines Development Group (GDG) at the European Commission Initiative on Breast Cancer, provides his personal views on some of the benefits of using AI in breast cancer screening programmes.

According to the World Cancer Research Fund, breast cancer is the most common cancer in women worldwide and the second most common cancer overall, with over two million new cases in 2018. As such, the detection of breast cancer is of paramount importance, with screening programmes established in an effort to find breast cancers early. However, the manual reading of mammograms is both time consuming and often leads to false positive/negative results. One solution to this is the development of Artificial Intelligence (AI) technologies which could assist radiologists/radiographers in the detection of breast cancer.

The Innovation Platform’s International Editor, Clifford Holt, spoke with Dr Axel Gräwingholt, Clinical Co-Chair of the Guidelines Development Group (GDG) at the European Commission Initiative on Breast Cancer, to discuss some of the benefits of using AI in breast cancer screening programmes.

AI is having an increasing impact across sectors, not least in healthcare. When it comes to the technology itself, how would you describe its development in recent years? What more needs to be done to ensure that the benefits are properly conferred to areas such as cancer screening etc.?

AI has undergone a tremendous development in recent years, especially in breast cancer detection. The older CAD-systems showed many false positive results and therefore had a low specificity. For radiologists, the workup of the many lesions was often annoying rather than helpful. The new generation with deep learning algorithms, however, is much more specific and also very sensitive. Studies by E. Connant (for 3D) and S. Köbrunner (for 2D) have shown this quite clearly.

To maintain qualification in mammography reading in screening programmes in countries such as Germany, for instance, we have to take annual or biannual tests, reading specially developed test sets, where our sensitivity and specificity is measured. It would be interesting to see the performance of such algorithms on the test sets, which would also be a good measure for the possibility of implementation of such an algorithm as a second reader. If the algorithm passes the test, it is likely that it would also perform well in the screening setting.

How can AI aid breast cancer screening specifically? What successes have already been seen? How is Europe faring in terms of leadership? Is there a sense that US-based tech giants such as Google etc. will dominate the arena?

AI could be used in different ways in this context. It can be used as a supportive tool for readers in screening programmes (which is done already in many cases), for instance. In the future, I would predict that there might be a possibility for AI to serve as a second reader, or perhaps as a decision aid for further recalls. On the other hand, it might be possible to have ‘true normal’ algorithms, which would be able to sort out the screening mammograms that do not even need to be seen by a radiologist.

The argument is not to replace radiologists, but rather to focus them on the more complicated cases where experience is needed and where examinations need to be compared with prior results etc. Of course, prior-comparison capabilities are also being developed by AI-companies, but issues such as different positioning can prove to be problematic.

Regarding industry, there are many companies working in the field of AI around the world, and, currently at least, there are no signs to suggest that a huge company such as Google would be the one to develop such specialised programs.

As part of the answer to the question of whether ‘an optimal number of readings vs. no specific number [should] be used for allowing mammography readers to work in mammography screening programmes?’, the ECIBC recommended more research into the role of artificial intelligence in double reading. Why was this recommended, and how would you like to see it being approached?

The research recommendation was partly based on the fact that in Europe and probably also in the rest of the world, we are facing a lack of breast-dedicated radiologists which are needed to maintain the high quality standards in breast cancer screening programmes. Thus, if there is a possibility to have experienced readers be supported by a suitable algorithm in reading screening mammograms, we might be able to overcome such shortages. Furthermore, an additional ‘reading’ through an algorithm could even outperform individual reading quality amongst radiologists. The development of this approach would require a trial – possibly a randomised controlled trial. And I dare to say that knowing the results generated from retrospective studies, I believe it would also be ethical to do so, as we are unlikely to miss cancers in such a set-up.

breast cancer detection
© iStock-Radovanovic96

Multiple studies have shown AI’s ability to surpass human experts in breast cancer prediction, and there is perhaps now a need for clinical trials to improve the accuracy and efficiency of breast cancer screening. How would you like to see this happen?

Breast cancer prediction is a very difficult subject. As such, perhaps newly-developed AI-risk tools could be best employed to provide even earlier detection of breast cancer than we are able to achieve at the moment; detections that we could not have made using just a mammography , but only through the addition of supplemental methods such as ultrasound, MRI, or contrast enhanced mammography (CESM) in women identified as being at higher risk.

We must also think about screening strategies based on measurables that could be applied – AI solutions could enable us to screen those women identified as high risk more frequently and/or with supplemental methods, and to screen those identified as being at low risk less frequently. But for all this, we need reliable measurements independent from individual perspectives, and these can only be provided by suitable algorithms. In addition, trials are potentially needed not only retrospectively but prospectively on large cohorts to find statistically difficult results.

Is there a need for radiographers etc. to already begin to receive training in AI technologies in anticipation of their wider roll-out?

Typically, working with these algorithms does not take much training – it takes just a few hours for a radiologist/radiographer to know how to use the detection software. The more interesting feature for radiographers, however, is that there are already developments in AI to assess the quality of images immediately after the images have been taken, to even tell them where quality problems are – for example incomplete imaging of the tissue in the inframammary fold etc., and to provide them with recommendations on how to improve this. This would minimise the technical recalls and also be a good tool for continuous education.

A recent report from Philips has found that before the COVID-19 pandemic, 60% of younger healthcare professionals ranked AI as the top digital health technology that would most improve their work satisfaction, with 39% identifying telehealth as the top technology. Yet, 61% of younger doctors now rank telehealth as the digital health technology that would have most improved their experiences at this time, with AI falling to 53%. Can you see any reasons for this?

The main focus for most healthcare professional given the COVID-19 pandemic is avoiding infections. Of course, teleradiology is how radiologists are able to work remotely, but even when doing teleradiology from home, having the colleague ‘AI’ next to you to help you with the cases would surely be beneficial.

I am sure that theses priorities will balance out once the pandemic is over or at least can be handled better. And, after all, telemedicine for radiologists is mostly about increasing the speed at which one can work – the transfer of images and reports, etc. AI, on the other hand, provides ‘medical help’, improving the quality of and satisfaction with your own work, and these algorithms are constantly improving.

Another problem with AI systems is the very narrow focus they tend to have. Usually, an algorithm focusses on one disease only – e.g. breast cancer detection – which leads to tremendous costs for a radiology institute if they want to use algorithms for several applications.

To return to the points made in the Philipps report: it is important to define what is meant by ‘satisfaction’. Is it referring to comfort and safety because of the ability to work from home (telemedicine)? Or is satisfaction also obtained from striving for better quality in patient care (as is possible with AI algorithms)? A shock like the coronavirus pandemic can really change people’s views – the first priority has been to maintain the medical standard even without patient contact. Once this can be provided, the focus will turn to the improvement of patient care, at which point AI will take the lead once more in terms of ‘satisfaction’.

Moving forwards, how do you predict the use of AI to aid detection of breast cancer in mammography will evolve?

In my personal opinion, I believe that AI will come to be used routinely in breast cancer screening and detection. It can help in many ways to increase early breast cancer detection either via detection algorithms or even earlier by ‘risk products’. The change from a one-size-fits-all to personalised screening approaches based on risk could be made in the future if validated risk AI-tools can be applied and are followed by recommendations for further strategies.

The quality of imaging could also be improved while the levels of stress experienced by radiologists who are required to read thousands of images could also be significantly reduced with the introduction of a program which is able to sort out the ‘normal’ results, thereby allowing the radiologist to focus on where his/her expertise and experience is in demand.

AI will not replace radiologists, but it will allow them to be more focussed and to improve quality.

Dr Axel Gräwingholt
Radiologist
Radiologie am Theater Paderborn
Clinical Co-chair, GDG, European Commission Initiative on Breast Cancer
axel.graewingholt@t-online.de
http://www.radiologie-am-theater.de/team.html

Please note, this article will also appear in the fourth edition of our new quarterly publication.



Subscribe to our newsletter

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Similar Articles

More from Innovation News Network