Advancing AI for healthcare is Christos Matsoukas’ call

Christos Matsoukas is developing robust AI models to grade chronic kidney disease. The models have reduced the evaluation time for pathologist by 97 percent and is now in production use. It all started with a WASP industrial PhD.

Christos Matsoukas never thought he would work with AI. Back in his home country Greece, when doing his bachelor, he was a physicist.

“I was working in astrobiology-related projects with people from NASA, and while doing that, I realized I was not really doing physics; I was doing computer vision.” Christos says.

While starting to explore the latest in machine learning, Christos heard rumours about a top-level machine learning master’s program at KTH Royal Institute of Technology in Sweden. Sweden had a good reputation in his eyes: the education system, the research community, as well as the work-life-balance. He decided to enrol, and, got accepted.

Best Thesis of the Year 2024

Christos master thesis’ supervisor, Hossein Azizpour, associate professor at KTH, recommended him to Kevin Smith and Magnus Söderberg for an industrial PhD position. Kevin is an associate professor in computer vision and biomedical image analysis at KTH, and Magnus is a pathologist and a senior director at AstraZeneca. The PhD position was a joint project between AstraZeneca and KTH, funded by WASP, and affiliated with SciLifeLab.

“I had already signed a contract with another Swedish company and suddenly got an e-mail form Kevin and Magnus saying ‘hey, just so you know, there is this new initiative from WASP, and we have a new project proposal, do you want to be a part of it?’”, Christos says.

First, he was not convinced. Doing a PhD was indeed a dream for him, but not an industrial PhD.  He feared that it would restrict his research, but, after many meetings with Kevin and Magnus during the summer of 2018, he decided to do it. And luckily, his fears were not answered:

“For me, that was definitely not the case. It went super great and smoothly, and I know this is mainly thanks to my supervisors.”.

For his thesis, he received the Best Thesis of the Year Award at the WASP Winter Conference 2025. Since the start of his PhD in 2018, Christos research has resulted in 13 publications, including high-impact venues such as CVPR, ICML, and NeurIPS. His work has a common theme: develop robust AI models that address data scarcity – a critical challenge in healthcare.

From left: WASP PhD students Johan Fredin Haslum, Christos Matsoukas, Joana Palés Huix, and associate professor Kevin Smith, at the WACV 2024 conference in Hawaii.

Technical innovations for medical image analysis

In his thesis Artificial Intelligence for Medical Image Analysis with Limited Data, Christos Matsoukas deals with enabling AI solutions for medical image analysis.

“AI is great, we all love it. However, most well-known models have been trained on data from across the entire internet. In contrast, medical image analysis, and medicine in general, suffers from limited data availability because of privacy and ethical concerns. So, how can we effectively use AI in medicine when these models require large amounts of data to perform well?”.

Christos thesis presents smart ways of mitigating these problems. One key is transfer learning. It is an approach where a model has been trained in one setting and then is used to improve generalization in another setting – you transfer the knowledge to another dataset.

“A challenge with transfer learning is the difference between the target domain and the source – which features can be used, and which cannot? This challenge is not unique to medical image analysis; for example, climate change forecasting also struggles with similar issues.”, says Christos, and continues:

“In my thesis, I explore why transfer learning works despite significant gaps and which features are effectively transferred between domains. This results in practical guidelines for efficiently leveraging foundation models and transfer learning in low-data tasks.”.

From Christos’ paper “What makes transfer learning work for medical images: Feature reuse & other factors”, appeared in CVPR 2024. The figure shows factors affecting the usefulness of transfer learning from the natural to medical domains. The size of each dot represents the observed gains from transfer learning and the color of the dot the reuse of features. The benefits from transfer learning increase with reduced data size, smaller distance between the source and target, and models with fewer inductive biases such as vision transformers.

AI reduced evaluation time by 97 percent

In addition to transfer learning, the other key spells self-supervised learning.

“Even if we successfully use transfer learning, we still require experts to annotate the data. These annotations are very expensive, difficult to obtain, and inherently biased. Therefore, we need to find ways to work around these annotation issues and streamline the training process”, says Christos.

At AstraZeneca, together with a pathology team, Christos developed an automated AI pipeline to grade chronic kidney disease – a disease that ten percent of the population suffers from. The pipelines have drastically reduced evaluation time for pathologists.

“We improved the efficiency by 97 per cent. Now, we can better utilize the pathologist’s time.”, Christos says.

As for the bias problem, Christos and his team relied on self-supervised learning. The results showed that the AI models performed better than the expert pathologist in the sense that the model agrees with each pathologist more than they agree with one another, and it maintains a consistent bias profile across different studies.

The AI models was trained on data from animals; therefore, the final step was to translate the findings to human biopsies that appear with different features due to biological and tissue acquisition differences.

“We managed to significantly mitigate the translation problem by developing a domain adaption method that reduced the translational gap by 45 per cent. This is a huge performance boost with significant implications for the future of translational medicine”, says Christos, and adds that he and his team at AstraZeneca continuously do quality controls on the tools.

Christos’ AI pipelines are now in production use globally, and helps experts make more robust evaluations every day.

AI is so much more than chat bots

After the dissertation in 2024, Christos stayed at KTH for almost a year as a research scientist and supervisor. At the beginning of 2025 he was recruited to AstraZeneca as associate director of AI. He will now continue working on AI solutions in medical image analysis but also focus on other domains within AI in healthcare.

“Advancing AI in healthcare can make expert-level diagnostic tools and treatments available for everyone, regardless of location or financial status. For me, that is the most meaningful aspect of applying AI to medicine.”, says Christos.

He is convinced that the Wallenberg Foundation, with WASP and DDLS at the forefront, is the enabler for world-leading, socially responsible AI research in Sweden.

“The Wallenberg Foundation give the opportunity to everyone – people from all different backgrounds. They enable AI for the good of humanity. It is important to highlight that AI can be used for so much more than just creating chat bots.”, Christos says.