Guides, project breakdowns, and lessons learned from building medical imaging platforms — written by the PYCAD team.
Hang a SWAN stack into a neuro MRI workflow: synthetic soft-ellipsoid CMBs on real negative hosts train a MONAI 3D U-Net that keeps ~84% Dice / ~88% lesion sensitivity vs real-trained when CMB labels are scarce. Hsu et al., arXiv:2610.00743.
Hang a PACS screenshot and CT stack into AnatomIQ: click the lesion, get matched cyan background ROIs for CNR, plus phantom-free NPS and TTF from patient anatomy. ~2-3 min/case. arXiv:2610.01686.
Hang incomplete T1w/T2w/FLAIR/DWI stacks into hetero-modal WMH+ISL segmentation with MMAR corruption routing. On mostly uni-modal training, hetero-modal mean macro DSC rises to 47.93% vs 42.29% specialists. arXiv:2610.00553.
Hang a thoracic planning CT into DAMM-Net++: multi-structure OAR and target overlays, calibrated uncertainty for triage, and TPS-ready RTSTRUCT. Macro Dice 0.955 / HD95 ~3.78 mm; reader study ~75-80% less contouring time. arXiv:2609.16036.
Hang a whole fetal ultrasound study into cardiac-gated MIL (FUSE CHD screen): case-level CHD vs healthy, optional critical vs non-critical severity, plus top attention-ranked frames for review. Internal AUC 0.985 / specificity 0.990; CARDIUM 0.513 to 0.944 with CORAL. arXiv:2609.31376.
Hang motion-corrupted DWI stacks into GraphSVR: stack-wise rigid slice-to-volume transforms that cut grid/rotation error ~73% vs FSL eddy under severe motion, strongest on sparse 12/24/32-direction jerky cases. arXiv:2609.24732.
We build custom medical imaging platforms — advanced DICOM viewers, AI segmentation, and the clinical systems around them.
Copyright © 2026 PYCAD. All Rights Reserved.