Guides, project breakdowns, and lessons learned from building medical imaging platforms — written by the PYCAD team.
Hang one low-dose CT measurement. PIPG-LDCT builds two physics-driven noisy training branches from that single acquisition, trains a Noise2Noise-style U-Net, and returns a cleaner stack without paired normal-dose labels. arXiv:2609.02654.
Hang a bone US frame, paint a rough brush on the interface. ExiL completes and refines that imperfect mask with a mask-conditioned U-Net and expert-mode updates. arXiv:2609.00473.
Hang a colonoscopy frame or prostate MRI slice, look at the structure. GazeRefine turns duration-weighted fixations into a dense overlay on frozen DINOv3. No labels, no fine-tuning. arXiv:2609.01310.
Untracked 2D ultrasound frame plus a 3D volume in. Rigid six-DoF pose out, no hardware tracker. State-space correspondences then closed-form Kabsch (SCoPE-Reg), IKIM Essen. arXiv:2608.28715.
Cardiac CT or MRI in. Seven chamber and vessel overlays out from one 3D nnU-Net, no modality flag. Label-to-image mix (LISynSeg), Vanderbilt/Mayo. arXiv:2608.31073.
Partial intraoperative lumbar US bone surface in. Watertight whole-spine mesh out in about 8 seconds. Zero-shot octree diffusion (ZODIAC), TUM CAMP. arXiv:2608.24422.
We build custom medical imaging platforms — advanced DICOM viewers, AI segmentation, and the clinical systems around them.
Copyright © 2026 PYCAD. All Rights Reserved.