PYCAD BLOG

Insights on Medical Imaging, AI & DICOM

Guides, project breakdowns, and lessons learned from building medical imaging platforms — written by the PYCAD team.

Magazine before/after from Fig. 4: two held-out brain tumor patients at 16x and 12x acceleration, each shown as the prior visit used as guide, SENSE, unguided LPDSNet, LGGS and the fully sampled scan, with zoom and error strips. Janjusevic, Chandarana, Wang, Feng, arXiv:2610.22643

LGGS: Faster Follow-Up Brain MRI Using the Patient’s Last Scan

LGGS rebuilds heavily undersampled follow-up brain MRI using the patient’s prior exam as a guide, without copying it into the new image. arXiv:2610.22643.

October 19, 2026

Magazine before/after from Fig. 3: two brain MRI test cases, a 64 mT FLAIR and a 3T T1, each shown as the lower-field input, the UniField++ virtual higher-field output and the real 3T or 7T scan. Lin, Wang, Yao, Li, Sun, Deng, Yuan, arXiv:2610.22982

UniField++: Making Low-Field Brain MRI Look Like 3T and 7T

UniField++ turns 64 mT brain MRI into virtual 3T and 3T into virtual 7T, with a vision-language planner that flags regions likely to go wrong. arXiv:2610.22982.

October 19, 2026

Method-strip banner from Fig. 1: chest CT volumes at several time points pass through a 3D RetinaNet detector, a ResNet50-SWS malignancy classifier, GradICON box transfer between visits and LesionLocator segmentation. Romoli et al., arXiv:2610.16548

NoduleFlow: Automatic Lung Nodule Masks Across Follow-Up CT

NoduleFlow chains four public, frozen models to detect, score and segment lung nodules across a patient’s follow-up CTs with no clicks. arXiv:2610.16548.

October 18, 2026

Magazine before/after from Fig. 4: a glioma case shown as the non-contrast T1 input, the MoFD synthetic T1ce without and with GAN loss, and the real gadolinium T1ce, with zoomed tumour views below. Kim, Kwon, Park, arXiv:2610.21703

MoFD: Synthetic Contrast-Enhanced Brain MRI From T1, T2 and FLAIR

MoFD predicts a contrast-enhanced T1 brain MRI from non-contrast T1, T2 and FLAIR with a 3D U-Net that mixes sequence features and their differences. arXiv:2610.21703.

October 18, 2026

Magazine before/after from Fig. 13: an input chest X-ray beside DeBoneDiT bone-suppressed outputs with 1 and 12 DiT blocks, and below each the residual map against the dual-energy soft-tissue image. Sun, Fan et al., arXiv:2610.19466

DeBoneDiT: Soft-Tissue Chest X-Rays Without Dual-Energy Hardware

DeBoneDiT turns one standard PA chest X-ray into a dual-energy style soft-tissue image in about 3.4 seconds. Apache-2.0 code and weights, arXiv:2610.19466.

October 17, 2026

Magazine before/after from Fig. 4: zoomed sagittal roots and axial crowns of a dental CBCT reconstructed from 20 projections with FDK, R2-Gaussian and AD-GS, beside the full-scan reference. Li, Xie et al., arXiv:2610.14963

AD-GS: Sharper Dental CBCT From Fewer Projections

AD-GS fits each dental CBCT scan with 3D Gaussians, steering them toward enamel and bone so 20 to 30 projections keep root and crown edges. arXiv:2610.14963, BMVC 2026.

October 17, 2026

We build custom medical imaging platforms — advanced DICOM viewers, AI segmentation, and the clinical systems around them.