Repository navigation
A typed decision model for computed tomography (CT).
RevCT identifies 18 chest-CT findings and returns calibrated scores as typed decisions. On CT-RATE’s corrected 3,002-volume validation cohort, it achieves 0.6154 macro-F1 and 0.8715 macro-AUROC, leading NVIDIA’s published 18-class macro-F1 comparison.
The complete COLIPRI-based model is approximately 147M parameters in one 589 MB safetensors checkpoint. Use Transformers AutoModel and AutoProcessor after importing the installed rev_ct package.
Fast preprocessing averages 2.856 s/scan on the 64-patient A100 80 GB sample, 7.63× faster than NV-Reason-CT’s measured 18-question protocol.
Wheel · Source distribution · Documentation · Weights.
Interactive demo · DOI: 10.57967/hf/10861.
Apache-2.0 covers RevCT code and learned heads. The bundled COLIPRI encoder retains its MIT licence.