TU Dresden

Genomic variant effect prediction from a bidirectional absorbing-diffusion model over human DNA, pretrained with entropy-guided span masking.

Whole-slide pathology embedding framework that ranks tiles, keeps only the 25 most informative, and encodes a slide in 2.27 seconds.

Radiology CT framework pairing lesion-driven contrastive slice embeddings with attention pooling to predict clinical endpoints without fine-tuning.

Slide-level histopathology encoder that contrastively aligns tile embeddings from several patch foundation models through a Mamba-2 aggregator.

Diffusion autoencoder framework that morphs H&E tiles until a pathology classifier flips, exposing the morphology behind the prediction.

SMILES transformer pretrained to predict 113 RDKit molecular descriptors, giving embeddings that carry physicochemical properties into ADMET models.

DNA language model of the human genome with a vocabulary learned by byte-pair encoding rather than fixed k-mers, fine-tuned for genome biology tasks.

Threshold-agnostic spot detection for 2D and 3D fluorescence microscopy, locating transcripts at subpixel accuracy via stereographic flow regression.

Nucleus instance segmentation for fluorescence microscopy and H&E histology, predicting a star-convex polygon per pixel to separate crowded nuclei.