Dr. Seyed-Ahmad Ahmadi, formerly a researcher at the Technical University of Munich and Ludwig-Maximilians University Munich, is now a senior solution architect with NVIDIA, where he explores not only novel machine learning architectures for medical data, but also how to put them into commercial and clinical use. Although he works with medical data in many different forms, he is particularly interested in graphical data, ranging from those at the level of individual patients to those covering large patient populations. He is also a co-organiser of this year's Workshop on Graphs in Biomedical Image Analysis, which for the first time also hosts two workshops from previous years on related topics as special tracks: the workshops on Topology- and Graph-Informed Imaging Informatics (TGI) and Hypergraph Computation for Medical Image Analysis (HGMIA). Dr. Seyed-Ahmad Ahmadi, formerly a researcher at the Technical University of Munich, is now a senior solution architect with NVIDIA where he researches not only novel machine learning architectures for medical data, but also how to put them into commercial and clinical use. Although he works with medical data in many different forms, he is particularly interested in graphical data, ranging from those at the level of individual patients to those covering large patient populations. He is also a co-organiser of this year's Workshop on Graphs in Biomedical Image Analysis.
It is a challenge every year, but it's a good one. We stick to a certain workshop format that has worked well so far, but we always add elements to keep it fresh. Our core focus is learning on graphs and higher topologies, but the field itself moves incredibly fast. That means novelty almost comes naturally - both in methods and applications. For example, in 2022 and 2023 we saw a lot of work in digital pathology, whole-slide imaging and brain connectomics and population modeling. In 2024, we saw works on shape modeling and multi-omics, and we had an outstanding paper on scene graphs in surgery. This year again, the papers will bring exciting new innovations and application areas.
We highlight these developments by inviting multiple keynotes, up to three or four, even if we are time-constrained by a half-day format as in 2022/2023. With this year's full-day program, we can really showcase how the field is evolving, and our repeat attendees appreciate that perspective.
Whatever we do to keep our workshop exciting, a guiding principle for us is to make sure that everyone has a voice. That includes newcomers, PhD students just starting out, and even Master's students. Wherever possible, we give every author an oral presentation with Q&A, followed by a poster session for deeper discussion.
It can be, yes. Last year in Marrakesh, we saw a peak in submissions and accepted papers, plus several keynotes that had already been confirmed. The room was absolutely packed. To make it work, we shortened presentation times to 7 to 8 minutes and kept a very tight schedule - just 1 or 2 questions per paper during the session, with more time for discussion at the posters. It was intense, but people really appreciated the energy. It made for a lively room and some great conversations.
We also experiment with new formats. Last year, for example, we ran a journal club before MICCAI. The idea was to exchange on research and bring the community together virtually before meeting in person. It was a lot of effort and we couldn't repeat it this year but we will try again going forward.
But really the most important novelty this year is the joint workshop concept between GRAIL,TGI and HGMIA, bringing together learning on graphs, hypergraphs, and higher-order topological domains. That's a really big step up for us, I am really happy about it, and I am grateful to our co-organizers for collaborating on this effort, especially the track chairs Xiaoling Hu and Chao Chen for TGI and Xiangmin Han for HGMIA. GRAIL was founded back in 2017, and I joined the organization team in 2022 because I loved the idea of building a community around this young field. The field has evolved a lot since then - it's not just about simple graphs anymore - and the fact that these three workshops are now joining forces reflects that beautifully.
Several things excite me. First, the MICCAI papers in recent years on multiplexed or hypergraph learning. A nice example is the MICCAI 2022 paper by Niharika D'Souza and colleagues from IBM, in Tanveer Syeda-Mahmood's group, on population graph learning in multimodal cohorts. That really resonated with me, because at the time my own academic research was gravitating more and more toward multimodal fusion. More complex graph structures allow us to model relationships between modalities and patients in much more flexible ways.
Second, the next step beyond hypergraphs is topology-informed learning. Last year's keynote by Mustafa Hajij at GRAIL was a bit of an eye-opener to me, and I'm sure for many others, too. This year, we're following up with his hands-on tutorial from graphs to higher order topological domains, using the open-source Python SDKs created in his lab. I'm very excited about this direction, and I hope that the hands-on nature of this session will help our students adopt these techniques quicker, and explore more application areas in the years going forward.
Third, there's molecular and genomic modelling - a huge application area for graph and topology learning that lies a bit outside the immediate radar of our MICCAI community. One innovation there is incorporating inductive priors like invariance and equivariance. For example, the biochemical properties of molecules or proteins shouldn't depend on their orientation in 3D space. Instead of relying on heavy data augmentation during training, you can build those symmetries directly into the network architecture. That's mathematically elegant and very powerful, and I think it opens doors to translation into more native MICCAI domains as well, like modelling 3D surfaces of organs or bones and their interactions.
And finally, another interesting field is this marriage between graphs and transformers, and where we are going with that. So far, we are often using them for segmentation. I did a lot of work in that area, too, and I am not saying it is a solved problem, but in many cases, we already know how to segment quite well. The question is what to do with that now, in order to move forward towards better patient care? To me, the most exciting avenue is to look at patients and cohorts more holistically. Patients aren't just images, their data is inherently multimodal. Using graph- and transformer-based foundation models, especially LLMs, to encode and fuse modalities is a really exciting step forward.
Not really. I see them as complementary.
LLMs are incredibly useful, but it's important not to overdo it. For example, in multimodal fusion there's a tendency to translate every modality into text and let the LLM combine everything into one structured report. That works, but we shouldn't forget that specialized models can often represent modalities and events much more efficiently in continuous embedding space rather than discrete tokens. That gives them more expressivity and allows us to design true end-to-end architectures, instead of constantly translating back and forth between token and embedding space.
Fundamentally, graphs allow us to model relationships in our data, whereas transformers allow us to discover and leverage previously unknown relationships via pairwise attention - but they come with a huge computational overhead. So it's not a matter of one replacing the other, it's about choosing the right tool for the right job.
That said, LLMs do bring a unique advantage - they let us "talk to" our models. They provide a layer of transparency and interpretability, even down to verbalizing intermediate outputs. That can be valuable for regulatory compliance and, most importantly, to give clinicians more confidence in the system. Of course, the irony is that LLMs themselves are not fully understood, and explaining their outputs is another challenge altogether. But let's just ignore that for now as that is a different area...
That's a different conference! [laughs]
But to finish that thought, specialist models like graphs are not going anywhere. They operate directly in continuous space and can be built into end-to-end architectures. Compared to large transformer models, they are often very compact, fast to train and infer, and they also require much less data. This holds true especially for graph models, which can perform remarkably well despite low parameter counts. Graph models in particular can achieve impressive performance with relatively few parameters, which makes them very attractive for real-time applications like signal processing in surgery - a major strength that we should not overlook. Finally, it is possible to integrate inductive priors into GNNs and other architectures, which would otherwise be impractical or near-impossible to model into LLM transformer models, think of e.g. the invariance and equivariance properties of graph-based molecule embedding models I mentioned earlier. Or think of physically informed GNNs - these are excellent frameworks for modeling real-time tissue deformation and digital twins in surgery, but this direction has not been really explored in our community yet.
So in the end, graph-models and LLMs, or transformers in a wider sense, they will coexist and I am not worried about one or the other taking over. And that's exactly what makes it exciting to follow where the field is heading with GRAIL, and now with our joint efforts with TGI and HGMIA. It's really motivating to see how this little niche we started with has grown into a vibrant community at MICCAI.
For me, MICCAI is a chance to recharge. Working in industry, I don't always get to immerse myself in academic research as much as I'd like. So it's fantastic to catch up on the latest developments in just one week.
It's a great environment for experts from industry, academia and medicine to exchange, to learn and to innovate. For example, I'm curious about the Start-Up Village at MICCAI this year. Seeing new ideas being pitched and the kind of feedback they get, that is very inspiring.
But honestly, one of the highlights is always talking with students and younger researchers. They come with fresh ideas, often with a bit of shyness, but their questions and perspectives are so valuable. Those conversations are fun and energising.
MICCAI is intense - definitely not a vacation, despite the travel to South Korea [laughs] - but it does feel like a retreat. You meet, you learn, you laugh, you hang out in the evenings. It's exhausting in the best possible way, and I'm really looking forward to all of it.