shenao1995/RadGS-Reg-RecM

RadGS-Reg RecM — best reconstruction, 2026-10-25

The current best foreground-ROI reconstruction checkpoint from the completed 200-epoch point-RegM alternating joint run. Selected at epoch 174 by maximum full-validation foreground 3D SSIM, among the saved validation-best candidates. This is a jointly fine-tuned RecM, superseding the projection-only release of 2026-10-22.

Code · Project · RegM · Full 431-case test dataset · Paper

Checkpoint and validation

best_detail.pth is the extracted RecM state dictionary, compatible with the accompanying config.json. It contains 255,879,088 parameters. Training completed 200 epochs; the saved epoch174 checkpoint maximizes validation foreground reconstruction SSIM. Model selection uses 431 validation vertebrae, not the test set. No checkpoint dominates every metric.

Metric	Full training validation, epoch174
Foreground ROI 3D SSIM	0.810497
Full-volume 3D SSIM	0.953858
Full-volume PSNR	24.0101 dB
Projection 2D SSIM	0.972758
Projection PSNR	33.3000 dB

These values are from the full distributed training validation, not a new independent reevaluation or test result. See evaluation.json, best_detail.json, release.json and manifest.json for exact numbers, selection, provenance and SHA256 hashes. Metric definitions and physical intensity windows are retained in the source snapshot; do not compare against older differently normalized metrics without matching protocols.

Download and use

from huggingface_hub import snapshot_download
snapshot_download('shenao1995/RadGS-Reg-RecM', local_dir='RecM')
snapshot_download('shenao1995/RadGS-Reg-Test', repo_type='dataset', local_dir='RadGS_test')

Extract source.tar.gz and use its gs_infer.py, or the current GitHub code:

python gs_infer.py --checkpoint /path/to/RecM/best_detail.pth \
  --data-fold /path/to/RadGS_test --output /path/to/predictions

Paired AP/lateral inputs are original views 25/0, area-resized to 128². The dataset loader applies the original scanner scene scale; model input scale is 1000 and predicted density is divided by 1000. Output: 128³ XYZ voxels, 1 mm, centered RAS frame. Gray webpage thumbnails are not native-intensity inputs.

For evaluated paired registration, use joint_best.pth from the RegM repository. That file contains the exact epoch168 RecM + RegM pair. Combining this separately selected epoch174 RecM with epoch168 RegM is possible but is not the pair behind the published registration metrics.

The original projection-only checkpoint remains available in Git history and tag projection-only-20261022; consumers needing the old weights should pin revision 5dcb5de2f4b980cee1c35b5a0cace6e75f1084bd.

License and scope

Original RadGS-Reg research model; see LICENSE_NOTICE.md and bundled upstream notices. This card does not grant a broader license. Training uses CTSpine1K-derived simulated cropped lumbar vertebrae. Cite RadGS-Reg (MICCAI 2025) and CTSpine1K; no performance claim is made for uncropped clinical radiographs. Source code is included at the exact completed training snapshot, without compiled binaries or credentials.