UPath-X2CT

Pretrained weights for Geometry-Aware Biplanar Chest CT Reconstruction with Residual Latent Diffusion. Code, fixed splits, training and evaluation.

Released model

The current release provides one complete model checkpoint: upath_x2ct_final.pt. It embeds the frozen normalized autoencoder, coarse conditioning network, validation-selected step-22,000 residual diffusion, training-only projection normalization, CT-to-HU scale/offset, and projector calibration. manifest.json records its SHA256, size and source provenance.

EMA weights are materialized as ordinary PyTorch state dictionaries, preserving all tensor values and dtypes. No optimizer state or training data is included. The code trains all three stages and exports the same single-file format. There is no separate coarse checkpoint or standalone coarse result table in the current release. Earlier revisions retain the previous release layout.

Use

Clone and install the linked GitHub repository, then run:

python scripts/download_weights.py --output weights
python scripts/predict.py --weights weights/upath_x2ct_final.pt --input projections.npz \
  --case-id example --output runs/example

Input is calibrated 512×512 AP and lateral CT-derived DRRs, with explicit camera matrices in centered LPS coordinates. Inference does not require target CT or masks. Output is a 128³ CT mean in HU plus sample standard deviation, using K=8, 50 DDIM steps, eta=0. See the code repository for exact input geometry.

Training and evaluation

Training: LIDC-IDRI, 817 patients / 824 CT volumes. Validation: 91 patients / 91 volumes. Held-out LIDC: 102 patients / 103 volumes. External tests: MSD Lung (63) and CT-RATE (425); neither was used for fitting. Seed: 20260724.

Historical final-model results (PSNR dB / SSIM percent / whole-volume MAE HU):

Cohort	PSNR	SSIM	MAE
LIDC-IDRI	26.35	81.45	50.00
MSD Lung	24.98	76.43	63.14
CT-RATE	24.14	71.84	70.99

These are patient-level means from the original evaluation. Release-time replay measurements and numerical drift are documented in the GitHub repository. Standalone coarse results could not be reconciled with the local replay; the author reports a precision-related reproduction issue, and some paper results were produced on other servers. The release checks have not established the cause. Bundling the internal components does not resolve this issue or promise exact reproduction of historical final-model predictions.

Intended use and limitations

Research on calibrated biplanar DRR-to-CT reconstruction. Performance on real clinical radiographs and diagnostic replacement are not established. Sampling standard deviation is not calibrated clinical confidence. No raw patient images or private clinical datasets are included.

The associated source code retains Apache-2.0. Dataset, ImageNet initialization and dependency terms remain separate; this card does not assign a new license to those upstream assets. See THIRD_PARTY_NOTICES.md in the code repository.