Guides, project breakdowns, and lessons learned from building medical imaging platforms — written by the PYCAD team.
FOE hangs the tumor mask, then flags silent segmentation failures from Sparse Autoencoder concepts inside MedSAM, without rewriting the segmenter. Prostate MRI, pancreatic CT, brain MRI. arXiv:2609.08879.
SpFiLM conditions a 3D U-Net with spatial FiLM so one network parcellates plain T1 and post-gadolinium T1ce into 106 brain regions without a second adaptation stage. arXiv:2609.07718.
IGRA anchors each chest-CT abnormality instance to its report finding during training, then drops every text component at inference. Hang the volume, get multi-label overlays for 14 ReXGroundingCT categories in one forward pass. arXiv:2609.00447.
CoMLP fuses head-and-neck PET/CT (and the clinical report) with cooperatively gated MLPs so modalities meet at full resolution before downsampling. Primary tumor and node overlays without pairwise attention matrices. arXiv:2609.04781.
Approximate box on a 29 MHz micro-US slice in. Frozen MedSAM plus CLAHE, dilation, and Fourier contour smoothing hang a training-free prostate overlay for PSAD or biopsy registration. Point clicks fail on speckle. arXiv:2608.14796.
Guangzhou-trained pediatric pneumonia AI keeps usable AUROC in Bangladesh and Vietnam, but the locked source threshold drops sensitivity to 6.2% and 0%. Recalibration restores alerts at a steep false-positive cost. arXiv:2609.05140.
We build custom medical imaging platforms — advanced DICOM viewers, AI segmentation, and the clinical systems around them.
Copyright © 2026 PYCAD. All Rights Reserved.