My main research focus over the past two years has been 3D foundation models for radiology, culminating in nnFoundation: a pair of complementary convolutional and transformer-based models pretrained on 2.1 million CT, MRI and PET volumes from 125 datasets, evaluated on 108 downstream tasks and released integrated into nnU-Net and nnDetection. I led the data infrastructure group, harmonizing over a petabyte of data and building the database, storage systems and data formats for training at that scale. On the pretraining side I worked on 3D DINOv2 and compared it against MAE under matched compute: DINOv2 was strong on classification, while MAE came out ahead overall, especially on segmentation. I also co-led the hands-on fine-tuning for segmentation and classification. I currently work with DINOv3, focusing on parameter-efficient fine-tuning (PEFT) across medical and non-medical image domains.
3D segmentation methods with an emphasis on weak supervision, interactive and prompt-based segmentation and usability-driven model design, including large-scale benchmarking, Napari-based interactive systems and scalable training and inference pipelines.
Collaboration with clinical partners on applied radiological segmentation tasks, translating methodological advances into clinically meaningful and validated systems.
Open-source tools for patch-based processing, large-volume data handling, compressed storage, experiment configuration and reproducible research workflows.