Pattern Recognition Lab

22/10/2026

Too Much to Look At: Teaching GUI Agents Where to Focus

Agents that operate phones and computers now receive screenshots, audio, video, their own action history and device state at every step. A new preprint from Nanyang Technological University and Alibaba shows how often that hurts: the most important input changes between almost half of adjacent steps, and nine in ten steps contain an input that makes the right action less likely.

Their small adapter, FocusGUI, roughly doubles action-type accuracy on a frozen model. The appendix shows that most of that comes from the trained injection of all channels; the per-step focusing adds a few points. A useful diagnosis, read with one benchmark and offline evaluation in mind.

Agents that operate smartphones and computers on our behalf no longer just look at a screenshot.

21/10/2026

A Regulator, Not an Optimizer: What the Network Around a Cortical Core Really Does

In layer 5 of the cortex, a small group of neurons is connected much more strongly than the rest. In our new paper in Frontiers in Computational Neuroscience, led by Claus Metzner, we asked whether the surrounding network is there to maximise the information flowing through such a core.

It is not. Without the periphery, the core saturates and its information flux collapses. With only constant mean inputs it does slightly better than with the full periphery, and a simple local rule that keeps each neuron active half the time does far better. The surrounding network keeps the core out of trouble, and that is all.

The limits are in the piece: binary model neurons, no task performance measured, and a hypothesis about cortex rather than a measurement.

In layer 5 of the cortex, a small number of neurons are connected far more strongly to each other than to the rest.

20/10/2026

The AI-BAY-HEALTH Hackathon is pushing the Bavarian Health Model ahead.

We had a great meetup at TU Munich with colleagues from LMU Munich, OTH Regensburg, UTN, FAU and Baiosphere. It is a real pleasure to see so many Bavarian institutions pulling in the same direction. Building medical AI for Bavaria works best together!

20/10/2026

Asking the Scan a Question: A Survival Model for PSMA Therapy, Read With Its Error Bars

Which patients will respond long-term to PSMA radioligand therapy for metastatic prostate cancer is still an open question. A new study in the Journal of Nuclear Medicine from TU München and Bern University Hospital lets a model answer it the way a physician would: baseline and interim scans, PSA and a clinical query, answered as 24-month survival.

The results are promising and honestly reported: 81.8 percent accuracy with a confidence interval from 0.52 to 0.95 on 110 patients, 71.9 percent externally after few-shot fine-tuning, and accuracy comparable to a model without the question. A proof of concept worth following, read with its error bars.

Radioligand therapy against the prostate-specific membrane antigen, PSMA, has become an established treatment for metastatic castration-resistant prostate cancer.

19/10/2026

Thirty-Five Interactive Systems, Simulated Users: Human-Centred Research at MICCAI

A MICCAI 2026 workshop paper counts how the community evaluates systems meant for human hands. Of 118 papers with human-centred themes, 35 propose interactive systems and test them only with simulated users.

An interactive system that has never met a real user has not been evaluated on the property that makes it interactive.

Medical imaging and computer-assisted intervention increasingly build systems that someone has to use: an interactive segmentation tool that takes a radiologist’s clicks, a navigation display in an operating room, an explanation shown next to a prediction.

17/10/2026

It was a pleasure to visit the Bavarian Landtag in Munich this week, together with Baiosphere and a delegation of the Shanghai Shenkang Hospital Development Center, which brings together 37 hospitals across Shanghai serving a population of 26 million people.

As everywhere in the world, digital processes are the fabric that modern medical AI is built on. Let's develop this together.

16/10/2026

Best of 25 Is Not a Result: How Layer Selection Inflates Speech Depression Detection

Paula Pérez-Toro from our Pattern Recognition Lab, with colleagues in Valencia and Munich, shows what picking the best of 25 encoder layers on your evaluation folds does: up to 0.09 AUC of optimism, and above-chance results even with shuffled labels.

Under honest nested selection, a compact interpretable model beats the deep encoders. A paper that makes other papers' numbers smaller, and the field better.

Pretrained speech encoders have become the default way to turn a recorded interview into features for clinical prediction.

15/10/2026

Physics on a Spring Budget: What a Differentiable Mass-Effect Solver Does and Does Not Recover

A brain tumour pushes healthy tissue aside, and modelling that usually takes a per-patient PDE inversion. A TUM-led team, with colleagues in Zurich and at FAU, replaces it with a differentiable spring lattice that runs about 100 times faster.

It beats the established physics solvers on 134 patients, and the authors also report the registration ceiling, which shows how much is still missing. Fast, physically plausible, honest about the gap.

A growing brain tumour does not only infiltrate.

14/10/2026

Where More Labels Stop Helping: A Surgical Foundation Model Maps Its Own Limits

The CAMMA group in Strasbourg built a surgical foundation model and then measured where it stops helping. Perception tasks saturate at a single labelled video. Reasoning tasks, including the Critical View of Safety, keep needing more labels.

A group publishing a result that undercuts the usual scaling story about its own model deserves attention. Anyone budgeting surgical annotation now has numbers instead of intuition.

Every surgical AI project runs into the same wall: expert annotation is scarce and expensive.

13/10/2026

Congratulations to Florian Kordon on passing his PhD defense!

A wonderful talk and a truly masterful presentation of his smart research ideas. Two of them, if you want to read along:

Deep Geometric Supervision Improves Spatial Generalization in Orthopedic Surgery Planning (MICCAI 2022): https://doi.org/10.1007/978-3-031-16449-1_59

A principled representation of elongated structures using heatmaps (Scientific Reports 2023): https://doi.org/10.1038/s41598-023-41221-2

A big thank you to Holger Kunze and Siemens Healthineers for the wonderful collaboration. Well done, Florian!