TinyUNet3D: 3D Segmentation in Under a Million Parameters
Glasgow distilled a 139M-parameter SegMamba-V2 into a 0.957M-parameter 3D U-Net. Distillation pays off mainly when labels are scarce.
November 10, 2026
Guides, project breakdowns, and lessons learned from building medical imaging platforms — written by the PYCAD team.
Glasgow distilled a 139M-parameter SegMamba-V2 into a 0.957M-parameter 3D U-Net. Distillation pays off mainly when labels are scarce.
DKFZ hashed 564,885 public CT and MRI volumes, found at least 118,487 duplicate pairs, and released MAP, a JSON profile for data lineage.
An open nnU-Net trained on 1,579 bpMRI cases from five cohorts, with a wavelet augmentation that scrambles each sequence.
Send a CBCT scan to a patient, referring dentist, or lab as a browser link. Add implant STLs, an optional password, and a 30-day to 1-year expiry.
Ostium labels for all 1000 ImageCAS scans, rooted centerline graphs with radii, and open localizers with sub-millimeter median error.
Fuses every smartphone photo of a patient with risk factors, then refers through a sensitivity threshold each clinic can calibrate.
We build custom medical imaging platforms — advanced DICOM viewers, AI segmentation, and the clinical systems around them.
Copyright © 2026 PYCAD. All Rights Reserved.