Guides, project breakdowns, and lessons learned from building medical imaging platforms — written by the PYCAD team.
T1c + T2-FLAIR in; independently trained 96³ and 64³ 3D U-Nets; weighted late fusion cuts false brain-met flags on BraTS-METS. Jaume, Wang, Warfield, arXiv:2609.10825.
Hang a sparse tracked ultrasound bone cloud. UBone3D physics-rectified flow matching returns a complete vertebra you can hang for scoliosis follow-up without another CT. Chen et al., arXiv:2609.11506.
Hang a portal-venous liver CT, drop a loose point or box (or none). SiGA/DiSECT spectral adapters on frozen SAM return colorectal liver-met overlays. Mojtahedi et al., arXiv:2609.11703.
XPos3R takes a live X-ray and a preoperative CT and regresses an intraoperative pose with one anatomy-specific pretrained model, then optional seconds of test-time optimization. Su et al., arXiv:2609.10733.
LeCor meta-learns case adapters so each clinician click on a planning CT updates the lung tumour overlay on untouched slices, not only the session memory. arXiv:2609.09477.
CoReFuse-Med filters resampling junk on sparse T2/FLAIR in shallow layers and rebalances deep fusion so multi-modal brain overlays stay usable when naive fusion would otherwise lose to T1-only. arXiv:2609.10261.
We build custom medical imaging platforms — advanced DICOM viewers, AI segmentation, and the clinical systems around them.
Copyright © 2026 PYCAD. All Rights Reserved.