shenao1995/RadGS-Reg-RegM

RadGS-Reg RegM — point residual registration, 2026-10-25

The validation-selected epoch168 checkpoint from the completed 200-epoch point-RegM alternating joint run. The custom point_residual34 architecture uses local residual point encoding with stage depths [3,4,6,3]; it is not the official PointNeXt model or a voxel ResNet.

Code · Project · RecM · Full test dataset · Paper

Files and model selection

• regm_best.pth: standalone RegM state plus constructor configuration, epoch and validation information. It loads with the current Reg3D_infer.py / checkpoint_model interface.
• joint_best.pth: unmodified original epoch168 joint checkpoint, containing its matched RecM and RegM, both configurations, and validation information. Use this for end-to-end inference and reproduction of the paired metrics.
• source.tar.gz: exact training-run source snapshot. config.json, best_validation.json, evaluation.json, val_cases.csv, release.json and manifest.json record the protocol, validation, checkpoint identity and file hashes.

Selection minimizes the existing maximum mean axis error plus the full-volume SSIM penalty. Epoch168 is the best saved checkpoint under that criterion; it was not selected by minimum mTRE. The final epoch200 has a slightly lower validation mTRE but a worse selected score. All selection uses validation only.

Full validation: 431 vertebrae

Independent single-GPU reevaluation of the matched epoch168 pair:

Metric	Value
Translation x / y / z MAE	0.3436 / 0.3789 / 0.3481 mm
Rotation x / y / z MAE	0.5041 / 0.4327 / 0.9042 degrees
Rotation geodesic error	1.2732 degrees
mTRE	0.8379 mm
mTRE < 2 mm	97.22%
Reconstruction foreground 3D SSIM	0.810430
Reconstruction full-volume 3D SSIM	0.953842

These are validation, not test, metrics. Vertebrae from the same CT are correlated. Mean errors do not imply every case satisfies the thresholds. Full case records are included; maximum validation mTRE is about 13.1 mm. Results use online RecM reconstruction, not the old frozen reconstruction caches.

Download and inference

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

Using the exact bundled source snapshot (tar -xzf source.tar.gz; cd source):

python joint_infer.py --checkpoint /path/to/RegM/joint_best.pth \
  --data-dir /path/to/RadGS_test --reg-data-dir /path/to/RadGS_test/registration \
  --split test --cases 5 --output-dir /path/to/results

For new NIfTI inputs with matching initial physical coordinates and attenuation units:

python Reg3D_infer.py --checkpoint /path/to/RegM/joint_best.pth \
  --moving /path/to/moving.nii.gz --fixed /path/to/reconstruction.nii.gz \
  --output-dir /path/to/registration

Inputs share the centered RAS-mm frame; the output matrix is moving-to-fixed, and is applied as a rigid physical resampling. Point extraction uses 2048 surface points per input at attenuation threshold .005. The trained input canvas is 192³ at 1 mm, not the old 96³/2 mm cache resolution. No independent cloud centering or per-image max normalization is used.

The standalone best-reconstruction RecM in its own repository is epoch174. Do not substitute it into the epoch168 pair and attribute these validation results to that untested combination.

License

See LICENSE_NOTICE.md and all upstream notices in source.tar.gz. Cite RadGS-Reg (MICCAI 2025) and CTSpine1K. This release grants no broader model/data license than the upstream terms. Research use on simulated vertebra-level data.