Pattern Recognition Lab

05/10/2026

Deadline extended! Due to popular demand, abstract submission for the 2nd Bavarian Conference on AI in Medicine is now open until 9 October.

If today's deadline came too fast for you: the abstract is only 300 words, so this is very doable this week.

BCAIM brings the medical AI community in Bavaria together on 30 November and 1 December at the Bavarian Academy of Sciences in Munich. I am organising it together with Katharina Breininger (Universität Würzburg), Björn Eskofier and Michael Ingrisch (LMU München), Heike Leutheuser (Universität Bayreuth), Dorit Merhof (Universität Regensburg), Julia Schnabel and Benedikt Wiestler (TU München) and Fabian Theis (Helmholtz München).

Looking forward to your submissions!

https://events.hifis.net/event/3582/

05/10/2026

Eleven Papers and Two Challenge Wins: What FAU Took to Strasbourg

MICCAI finished on Wednesday. The 29th International Conference on Medical Image Computing and Computer Assisted Intervention ran from 27 September to 1 October in Strasbourg, and accepted 1,167 papers out of 4,402 that went to review.

Erlangen brought eleven of them: six from the Pattern Recognition Lab and five from Bernhard Kainz's IDEA Lab. Add two challenge results — first place in PENGWIN's automatic fracture segmentation task, and three task wins plus a second place for Team Beyond Normals in VLM3D — and the RIME workshop, which we co-organise with Tübingen, Imperial College London and TU München.

Both challenge entries were joint work with Siemens Healthineers, and the VLM3D team also included d.hip. That part is not decoration: universities are good at methods and weak on data agreements, annotation at scale and keeping a pipeline alive after the doctoral student who built it has graduated. Partners are good at exactly that.

The write-up is a map rather than a review — I did not sit through all eleven talks, and I am not going to summarise results I did not see.

MICCAI finished on Wednesday.

02/10/2026

Both Columns, One Vendor: What the Opus 5.5 Benchmarks Can and Cannot Show

Yesterday I wrote about a mechanism: that the quality of a hosted model can be set administratively rather than technically, by a control the seller holds and the buyer cannot read. Today I want to look at a document.

Anthropic's launch page for Claude Opus 5.5 compares it against Opus 5 across nine benchmarks. It states that its own testing ran at max effort. It does not state the effort level used for the Opus 5 column — and the gap between two effort settings of the same model is exactly the quantity being advertised.

The prices did fall, substantially, and I say so in the piece: input from USD 5 to 4, cache reads cut sixty percent. Anyone arguing the old model is held back to upsell the new one should stop. But the same page says forty percent less to run — that is the cost of serving, not the price of buying, and the lever that cuts it is the one that decides how long the model thinks.

I am not alleging fraud. I am saying that no instrument exists which could tell us either way, that building one would be cheap, and that nobody has.

Yesterday I wrote about a mechanism.

01/10/2026

Congratulations to Team Beyond Normals — three task wins at the VLM3D challenge at MICCAI 2026, plus second place in MRI Abnormality Classification!

VLM3D is the Vision-Language Modeling in 3D Medical Imaging challenge, covering chest CT and brain MRI across clinically grounded tasks: report generation, abnormality classification, localisation and segmentation.

This one is a genuine joint effort — the Pattern Recognition Lab at FAU Erlangen-Nürnberg, Bernhard Kainz's IDEA Lab, Siemens Healthineers and d.hip. Four groups, one team, and it clearly worked.

Very well done, all of you!

01/10/2026

A fantastic evening with Prof. Jang-Hwan Choi and the team from Ewha Womans University on the last day of MICCAI 2026.

What makes this one special is how far back it goes. Our groups have been friends since Jang-Hwan and I overlapped at Stanford University many years ago. Two labs on opposite sides of the world, and the students still keep ending up at the same dinner table.

That is the quiet part of academic life that never makes it into a grant report, and it may be the most durable thing we build. Thank you for a wonderful evening, and to everyone from FAU Erlangen-Nürnberg and Ewha who made it such good company.

Until the next one!

01/10/2026

Turning AI Into Light Bulbs: Planned Obsolescence Does Not Need a Broken Part

In January 1925 the largest lamp manufacturers in the world met in Geneva and founded a company called Phoebus. Osram, Philips, the Compagnie des Lampes, Associated Electrical Industries, Tungsram and, through its international arm, General Electric agreed to divide the world market between them. The best remembered clause of that agreement fixed the service life of an ordinary incandescent bulb at one thousand hours. Longer-lived lamps had been sold and advertised before. The cartel ran a testing regime in which members submitted samples, and a member whose lamps lasted too long paid a penalty on a sliding scale. Not defects were fined. Durability was.

In January 1925 the largest lamp manufacturers in the world met in Geneva and founded a company called Phoebus.

30/09/2026

The Pattern Recognition Lab at MICCAI 2026 — present and former members, all in one photo.

This is the part of a conference that never shows up in the proceedings: people who have long since moved on to other labs and companies, standing in front of the same letters as the ones who joined most recently. The papers are why we travel. This is why we come back.

Great to meet up with the pattern recognition family again. Thanks to everyone who came by — and yes, the FAU Erlangen-Nürnberg hoodies do make us easy to spot in a crowd.

30/09/2026

The Midas Touch for Code: Scaling AI Coding Environments Straight from Source

Software-building agents have moved from clever autocomplete tricks to genuine “coding companions” that can write functions, fix bugs, and even drive a terminal session. The next hurdle for these agents is learning in an environment that resembles real development work: a steady stream of diverse tasks, each paired with a trustworthy way to tell whether the solution is correct. All the existing pipelines – from SWE-bench to SWE-smith – harvest tasks from development artifacts such as issue tickets, pull-request discussions, or existing test suites. While useful, those pipelines are limited by the coverage and quality of the artifacts they rely on.

Software-building agents have moved from clever autocomplete tricks to genuine “coding companions” that can write functions, fix bugs, and even drive a terminal session.

29/09/2026

PCCT Look, EICT Information: What SUMI Does and Does Not Distil

Photon-counting computed tomography (PCCT) promises sharper pictures of the lungs, lower noise, and the added bonus of spectral data that can help doctors spot tiny airways or early tumours. The hardware, however, is still a rarity—costly detectors sit only in elite research hospitals, leaving the majority of patients with conventional energy-integrating CT (EICT) machines that blur the finest details. A 2026 pre-print on arXiv tackles this inequality head-on. Its authors ask whether the visual advantages of PCCT can be “distilled” into routine scans using artificial intelligence, without ever having to acquire paired PCCT-EICT images from the same patient.

Photon-counting computed tomography (PCCT) promises sharper pictures of the lungs, lower noise, and the added bonus of spectral data that can help doctors spot tiny airways or early tumours.

28/09/2026

The Pattern Recognition Lab on the big stage at MICCAI 2026 — congratulations to Daiqi Liu for a brilliant talk!

The work is "Speech-Guided Multimodal Learning for Vocal Tract Segmentation in Real-Time MRI": segmenting tongue, velum, upper and lower lip frame by frame in real-time MRI, guided by the speech signal itself.

Real-time MRI is the only non-invasive modality that shows all articulators at once during continuous natural speech, and tracing their contours frame by frame is what makes the vocal tract quantitatively measurable. That matters clinically: after a glossectomy patients compensate with new articulatory strategies, and in Parkinson's disease articulatory coordination degrades early.

Very proud to see FAU Erlangen-Nürnberg represented like this. Well done, Daiqi!