Using AI in the fight against breast cancer

Mit KI gegen Brustkrebs

Each year, over 70,000 cases of breast cancer are diagnosed in Germany; worldwide, it was around 22.3 million in 2020. Breast cancer is still the type of cancer that causes the most deaths in women. Within a European consortium, Johannes Gregori, Professor of Physics and Machine Vision at Darmstadt University of Applied Sciences, and his doctoral student Yaqeen Ali are conducting research into new diagnostic methods: “BosomShield” is a large-scale project that aims to combine diagnostic procedures such as ultrasound, mammogram and biopsy and analyse them within an overall system supported by AI. This aims at enabling more accurate diagnoses – and more effective therapies.

By Christina Janssen, 18.9.2023

The puzzle that Johannes Gregori and his doctoral student Yaqeen Ali are endeavouring to put together is highly complex: together with colleagues from throughout the whole of Europe, they are working to create a new basis for breast cancer diagnostics. Ultrasound, mammograms, MRI scans, biopsies, genotyping – to date, the results of all these individual examinations have been looked at and evaluated separately. A flaw, says physicist Gregori, who has been teaching and conducting research at h_da for two years. “There is a gap between what we diagnose and what we could achieve through therapy. We want to close this gap with the BosomShield project.” The aim is to combine different diagnostic techniques and by doing so produce more accurate diagnoses. More precisely, the researchers hope that this will deliver more exact insights into tumour subtypes, probability of relapse and possible therapies. In this way, the project could contribute to increasing the chances of survival for breast cancer patients.

EU-funded doctoral network

Eight universities and two industrial partners in Germany, France, Italy, the Netherlands, Sweden, Slovenia, Spain and Poland are participating in the BosomShield project. Project coordinator is the Spanish University Rovira i Virgili in Tarragona. The European Union is funding the project within the Marie Skłodowska-Curie Doctoral Networks programme, which finances one doctoral student at each of the ten partner organisations. Since the young researchers are not allowed to come from the country in which they are employed, it was necessary to advertise each position internationally. Yaqeen Ali, Gregori’s doctoral student, hails from Pakistan, completed a Master’s degree in computer science at the COMSATS University in Lahore and spotted the call for applications on an international website for doctoral researchers. “At the kick-off meeting, I met all the other doctoral students in the project,” he reports. "Most of us are computer scientists, some are mathematicians, we all have strong links to machine learning and deep learning, which allows us to exchange ideas on the same professional level and to support each other.” 

Each partner is working on a specific aspect within the project. “For example, one colleague is working on relapse prediction, that is, the probability of a cancer recurring,” explains Ali. “Others are concentrating on evaluating mammograms, others on histological images, and so on.” On the German side, mediri is on board, a company based in Heidelberg that specialises in medical imaging, of which Gregori was managing director for ten years before joining h_da. Yaqeen Ali is employed at the Heidelberg company, and as a doctoral student he is supervised at h_da.  The partner organisations deliver individual pieces of the puzzle, and Gregori, Ali and other team members in Darmstadt and Heidelberg piece them all together, all the while observing the strictest data protection rules: “We are working on a computer-aided diagnosis system (CAD) in which all the datasets for a patient can be uploaded, combined and evaluated,” reports Professor Gregori. “Methods to do this already exist, but our project draws together all data and evaluation techniques in a cloud-based platform for the first time.”

Data protection: the No. 1 priority

For the AI system to function reliably later on, it has to be trained – with tens of thousands of tumour images previously classified by hand. In this way, the system is fed with information during the learning phase as to whether an image shows a benign or a malignant tumour. This is how the AI learns to recognise patterns. For this training, the h_da team uses image material from publicly accessible databases. For data protection reasons, it is not possible to use current patient data in this early development phase. “We therefore simulate the various hospitals in our model and assign the data to them,” says Gregori, describing the approach.

Protecting extremely sensitive patient data should also continue later on when the system is in everyday use. To solve this dilemma, a new approach comes into play in the BosomShield project: federated learning. Here, the individual organisations do not exchange the data they have collected, but instead training is done separately and on site with local datasets. “Only afterwards are the results amalgamated.” The principle here is as follows: If the mountain won’t come to Muhammad, Muhammad must go to the mountain. “Very few papers on this technique have yet been published in the area of breast cancer. That is the very core of the scientific work we are tackling within this doctoral network.”

Accurate classification of tumours with the help of artificial intelligence

One of the biggest challenges in the project is the fact that each hospital works a little differently – with other MRI or ultrasound equipment that is configured differently. The way the data are recorded can also vary. Figuring out how processing such heterogeneous data influences the results is therefore one of the key issues: “What happens when we combine images from different sources? Which errors can occur as a result?” At the end of the day, explains Yaqeen Ali, the system must recognise and offset differences so that all the “pieces of the puzzle” fit together and patients receive the type of therapy that is best for them personally.

Some doctors are sceptical about this development. “Deep learning systems are a black box,” explains Professor Gregori. “It is often unclear how exactly a deep learning system makes its decisions. We have to take this into consideration when developing and using them.” This does not make it easier for doctors to instill trust in such systems. In addition, there is also the question of whether artificial intelligence will at some point make radiologists superfluous. “Definitely not,” says Gregori. “But the job description might change radically. Radiologists will have AI-supported helpers to hand and can then do their job even more efficiently.”

“Help so many people”

Gregori, who has been working at the interface between physics, computer science and medicine since his thesis, has also embedded the topic of medical image processing in teaching at Darmstadt University of Applied Sciences. He has introduced a mandatory elective module on the topic for students of the Optical Technology and Image Processing degree programme, and in this way is contributing to training experts that industry is desperately looking for. “There are many areas where our students are in demand,” explains Gregori, “medical technology companies such as Sirona, which is now focusing on AI applications, or the optics company Leica, whose histology products, that is, apparatus for examining tissue, are also helpful in breast cancer diagnostics, urgently need specialists. And our graduates are naturally in demand in all the major software companies.”

Doctoral researcher Yaqeen Ali is also aware of this and has his professional future firmly in view: he would like to complete his doctoral degree in three years and then spend some more time working in Germany in order to acquire more in-depth knowledge. And perhaps also to get to know Germany and Europe a little better than his tough workload currently allows. At least the Doctoral Network programme includes four research stays of one to two months with various European partners as well as annual doctoral training courses, most recently this summer in Trieste, Italy. However, Ali admits that all this leaves little time for hobbies and friends. But that’s not a problem, he says: “We are doing something that can help so many people. That’s very motivating for me.” The most important thing is that all the pieces of the puzzle fit together.


Christina Janssen
Science Editor
University Communication
Tel.: +49.6151.16-30112

Translation: Sharon Oranski