Baseline NCCT In, Infarct Overlay and Net Water Uptake Out Without CTP or DWI
Hang an infarct overlay on baseline NCCT alone and read NWU without CTP or DWI for delineation. Fine-tune when sites shift. Britt et al., arXiv:2609.20151.
Ultrasound In, Mask Class Box and Biometry Out From One Frequency-Routed Stack
Hang one FreqDINO++ stack so breast, thyroid, and fetal US get mask, class, box, and biometry without four specialist models. Xu et al., arXiv:2609.20340.
Broken Ultrasound Lesion Overlay In, Progressive Failure Repair Out
Frozen ultrasound lesion segmentor in, ordered gross-miss then FN then FP repair out. FAPR rebuilds the broken overlay without retraining Stage-1. Wang et al., arXiv:2609.18256.
Congenital Heart CT In, Chamber Overlays That Survive Held-Out Site Shift
Hang CHD chamber and vessel overlays that still look like chambers on a held-out CT cohort. In-distribution Dice can hide the break. Vijesh et al., arXiv:2609.17068.
Echo Cine In, EF Out: Why LV Mask Overlays Do Not Lift Learned Ejection Fraction
Hang A4C echo cine and get EF from raw video. Explicit LV mask channels sit above the segmentation ceiling and do not beat direct regression. Khouzani et al., arXiv:2609.19730.
Point-Prompt Ultrasound In, Contours That Survive Device and Site Shift
SonoBase hangs interactive ultrasound overlays from one foundation pretrained on SonoCorpus, then derives EF and fetal biometry that hold under device and site shift. Qin et al., arXiv:2609.19230.
Scribble One Slice, Propagate Adapted Volume Masks With Fewer Re-Prompts
IMVS couples an online-adapted 2D Slice Mask Adapter with a frozen Volume Mask Tracker so each scribble updates the model for the slices ahead across 8 CT/MRI datasets. Reen, Borkar, Mahapatra, arXiv:2609.16775.
Pediatric Brain MRI In, Four Tumor Subregion Overlays Out for Planning
One compact 3D NeuroTS-Net hangs enhancing, non-enhancing, cyst, and edema overlays on BraTS-PEDs mpMRI without external data or pretrained weights. Peteleaza et al., arXiv:2609.16873.
CT and MRI Pancreas Masks From One Model, Head-Body-Tail After MRI-Only Labels
Shared nnU-Net plus CT-MRI domain alignment on 4,604 scans; freeze the encoder and train head-body-tail from MRI-only subregion labels, then hang the same overlay language on CT. Hong et al., arXiv:2609.13043.
Text Prompt In, ROI-Guided Cardiac Volume View Out
MedVA turns clinical text into an ROI intent graph, pulls TotalSegmentator and VoxTell masks on the source volume, then adjusts opacity with a volume visibility score. An et al., arXiv:2609.14874.