shenao1995/RadGS-Reg-Test-10

RadGS-Reg: 10 RecM test samples

A 10-vertebra demonstration subset of the 431-case RadGS-Reg RecM test split. Model · Code · Paper

These are the same ten cases used for the released qualitative figures, selected at evenly spaced original test indices: 0, 48, 96, 143, 191, 239, 287, 334, 382, 430. They were selected before predictions, without performance-based filtering. There is one vertebra from each of 10 distinct CTs. This is not the full benchmark.

Contents

Each sample has an original .pickle, accompanying .json geometry/provenance, and .ply 10,000-point cloud, compatible with the current RadGS-Reg loader. The pickle contains a centered 128³ float32 attenuation volume, 50 simulated 256×256 projections, view angles and scanner parameters. Native volume axes: XYZ, centered RAS; spacing: 1 mm. Projections span [0°,180°), excluding the endpoint. Input AP/lateral views are indices 25/0, resized to 128×128 by the model loader. Projection storage units follow metadata: line integral /1000; volume attenuation is in 1/mm. Keep all files for a sample together.

splits/test.txt enumerates exactly the ten sample filenames. selection.json records original test indices and sample identifiers. metrics_10_cases.csv contains per-case results from the released model. previews/ provides GT / prediction / absolute-error projection and volume figures, with a common GT-derived display window. Source JSON paths are shortened to basenames; source file mtimes are omitted, while reconstruction geometry is preserved.

Download

from huggingface_hub import snapshot_download
root = snapshot_download("shenao1995/RadGS-Reg-Test-10", repo_type="dataset", local_dir="RecM_test10")

Use gs_infer.py --checkpoint <model>/best_detail.pth --data-fold RecM_test10 --output predictions with the source/environment provided in the model repository. The original binary format is provided for compatibility; this release does not provide a datasets.load_dataset() builder or an automatic volumetric browser viewer.

Preview

Projection GT / prediction / error

Volume slices GT / prediction / error

Provenance and licensing

Derived from CTSpine1K (original project), using NanoDRR to create projections. Sources in this subset: COLONOG (9), MSD-T10 (1). The source card states CC-BY-NC-SA without a license version; this release preserves its attribution, noncommercial and share-alike terms. See LICENSE.md and the preserved SOURCE_CTSpine1K_README.md for terms and the CTSpine1K citation. Please also cite RadGS-Reg when using its reconstruction model and generated samples. These are simulated vertebra-level research samples.