Self-Training Coronary Vessel Masks With Pre-Contrast Frames
An nnU-Net labels a new hospital’s angiograms, and the frame before contrast arrives flags the branches it missed.
RepConTUR: Report-Prompted 3D Tumor Contours for Abdominal CT
The report names the liver segment or pancreas region, and a 3D model contours that tumor with fewer stray masks elsewhere.
Pancreatic Duct Width From 3D MRCP, With a Review Flag
A two-stage 3D model segments the main pancreatic duct on MRCP, and ensemble disagreement marks the scans a radiologist should check.
Plausible but Unimplantable: Foundation Models vs Cranial Implants
An audit of three released shape completion models on 239 skull defects found no output that could be placed in a patient.
SGAV-MP: Smaller Segmentation Models That Keep Their Uncertainty
A post-training mixed-precision method that checks calibration and uncertainty maps, not only Dice, before it quantizes a layer.
DART: One Shape Model That Repairs Broken Segmentation Masks
A text-conditioned 3D diffusion model cleans up masks from TotalSegmentator or MedSAM, on CT and MRI, without reading the image.
Hippocampal Segmentation Ensembles That Flag Their Own Failures
Averaging InnerEye, FastSurfer and HippoSeg gives calibrated hippocampus masks plus an entropy score that sends doubtful cases to review.
CAP-DIR: Adaptive Radiotherapy Registration That Reads the Plan
Uses planning contours, dose, case notes and earlier fractions to carry contours onto repeat CTs. No code yet; weights stay private.
ReSeg: Organ Masks on CT, MRI and PET Without Real Training Scans
Trained only on synthetic volumes built from organ masks, it outlines liver, spleen and kidneys on CT, MRI and PET. MIT code, no weights.
One Click on One CT Slice In, a Full 3D Organ Mask Out That Is Carried Slice by Slice From the Neighbours and Still Takes Class Prompts, Reference Images and Corrections in the Same Model
Click once on one slice of a CT or MR volume and UniPro carries the mask through the other slices, using the seed slice and the latest predictions as references. The same 2D model also takes a class name, a reference example or corrections. It is best of the compared models on 7 of 8 single-click tests, trails 3D nnU-Net in plain semantic mode, needs 28 seconds per abdominal volume, and the repository README still says “To be released”. Guo et al., arXiv:2610.06938.