Three Papers presented at MICCAI COMPAYL 2027 in Strassbourg

Sweta Banerjee has the honor of giving an oral presentation about her recent paper "Beyond Classification: Pathology Foundation Models as Detection Encoders for Mitotic Figures". In the talk, she will be talking about why foundation models are (despite their size) actually really interesting candidates as feature extractors for mitotic figure detection. While the in-domain performance was on par with SOTA methods, out of domain is where the model using the foundation model feature extracor really shines. The project was initially based on a Master's thesis in our group by Alireza Teimoury that has been extended into a full paper at COMPAYL by Sweta. 

Niklas Hargarter has another interesting story to tell: There have been a bunch of methods proposed for deep learning-based autofocus in microscopy. Many of those use specific input transforms and claim that the generalization significantly improves with these. Niklas found that - while this is true in many cases - standard augmentation yields just as good results. But still, there is quite a gap between in-domain and out-of-domain performance, as his paper "Investigation of Factors Contributing to Domain Generalization in Single-Shot Autofocus" shows.

Last, not least, we have a fascinating work by Marvin Carstensen to share: Do you also have an old, fully analog, light microscope sitting around collecting dust and wondered, how you could convert it into a digital setup? Marvin got you covered: In his work, "RetroScope: An Open Framework for Motorizing and Digitizing Vintage Microscopes", he presents a parametric 3D adapter and software framework that can be used to turn almost every old analog microscope into a digital one, including touch screen. We have the working prototype in our lab and it always amazes guests. Great job, Marvin!