Open Source
Research code I develop with the Medical Image Computing group at DKFZ.
LongiSeg
One framework for segmentation across timepoints. LongiSeg brings longitudinal context into nnU-Net and houses our work on temporal feature merging and on clinician-verified interactive lesion tracking. Fully compatible with nnU-Net.
LongiTrack Viewer
The full verified tracking workflow in a viewer. Click a lesion in the baseline scan, check the point that uniGradICON registration proposes in the follow-up, and LongiSeg segments the lesion in both scans.
LongiTrack Backend
The GPU service behind the LongiTrack Viewer, running registration and segmentation locally or on a remote server. It can be implemented against with any medical imaging viewer, such as MITK or 3D Slicer.
Skeleton Recall Loss
A loss that keeps vessels, nerves and other thin tubular structures connected. It computes a tubed skeleton of the ground truth during data loading and adds a soft recall term on it, which avoids the cost of differentiable skeletonization.
Temporal Difference Weighting
A modular block that merges features from two timepoints. It computes normalized attention maps from the differences between their feature representations, so the network focuses on regions that changed. It works as a drop-in block for UNet-style backbones.