Benchmark Dataset · Clinical CT to Micro-PCCT

LumbarSR

A paired clinical CT and photon-counting Micro-PCCT benchmark for lumbar vertebral super-resolution, released with aligned BoneMask ROIs, reference registration, visualization assets, and public benchmark tables.

Overview

LumbarSR is a paired clinical CT and Micro-PCCT benchmark for lumbar vertebral super-resolution, containing 30 ex vivo human lumbar samples scanned with multiple clinical CT configurations and matched 105 μm Micro-PCCT.

The task is to reconstruct bone microstructure from routine clinical CT into Micro-PCCT-like resolution for musculoskeletal imaging research.

Multi-resolution CT Visualization

Tri-planar views and 3D bone rendering across different CT resolutions
Click to view at higher resolution (Original ultra-high resolution is not available online)

Task Definition

Input: clinical CT with soft tissue kernel at two fields of view. Output: 105 μm isotropic Micro-PCCT.

Field of View In-plane Resolution Slice Thickness
Small FOV (High Resolution) 195 μm 500 μm / 1000 μm
Large FOV (Standard) 586 μm 500 μm / 1000 μm
Pipeline Flowchart

Pipeline overview (Full resolution)

Super-Resolution Visualization

The following figure shows a representative tri-planar comparison for Lumbar_26 using 586X_586Y_1000Z_S as input. Columns from left to right are Clinical CT, Nearest, SwinIR, SRCNN, UNet, ESRGAN, and the Micro-PCCT target.

Representative super-resolution comparison on Lumbar_26 at 586um FOV

Open the PDF version for the full-resolution figure.

Dataset

Download

Download Dataset from Zenodo (ZIP parts) DOI: 10.5281/zenodo.19404387

Zenodo DOI badge

The dataset is distributed on Zenodo as split ZIP archives. Please use the Zenodo record for the current file list and download instructions.

Directory Structure

LumbarSR/
├── original_dicom/              # Original DICOM data
│   ├── Lumbar_01/
│   │   ├── clinical_ct/
│   │   │   ├── 195X_195Y_500Z_S/      # Small FOV, soft kernel
│   │   │   ├── 195X_195Y_1000Z_S/     # Small FOV, soft kernel
│   │   │   ├── 586X_586Y_500Z_S/      # Large FOV, soft kernel
│   │   │   └── 586X_586Y_1000Z_S/     # Large FOV, soft kernel
│   │   └── micro_ct/
│   │       └── *.dcm                   # Original Micro-PCCT DICOM
│   ├── Lumbar_02/
│   │   └── ...
│   └── Lumbar_30/
│       └── ...
├── RegisteredData/              # Processed NIfTI data (ready to use)
│   ├── Lumbar_01/
│   │   ├── Lumbar01_ClinicalCT_195X_195Y_500Z_S_registered.nii.gz
│   │   ├── Lumbar01_ClinicalCT_195X_195Y_1000Z_S_registered.nii.gz
│   │   ├── Lumbar01_ClinicalCT_586X_586Y_500Z_S_registered.nii.gz
│   │   ├── Lumbar01_ClinicalCT_586X_586Y_1000Z_S_registered.nii.gz
│   │   └── Lumbar01_MicroPCCT_105um.nii.gz                          # Ground truth
│   ├── Lumbar_02/
│   │   └── ...
│   └── Lumbar_30/
│       └── ...
├── BoneMask/                    # Released binary bone ROI aligned to RegisteredData
│   ├── Lumbar_01/
│   │   ├── Lumbar01_MicroPCCT_105um_BoneMask.nii.gz
│   │   ├── Lumbar01_ClinicalCT_195X_195Y_500Z_B_registered_BoneMask.nii.gz
│   │   ├── Lumbar01_ClinicalCT_195X_195Y_500Z_S_registered_BoneMask.nii.gz
│   │   ├── Lumbar01_ClinicalCT_195X_195Y_1000Z_B_registered_BoneMask.nii.gz
│   │   ├── Lumbar01_ClinicalCT_195X_195Y_1000Z_S_registered_BoneMask.nii.gz
│   │   ├── Lumbar01_ClinicalCT_586X_586Y_500Z_B_registered_BoneMask.nii.gz
│   │   ├── Lumbar01_ClinicalCT_586X_586Y_500Z_S_registered_BoneMask.nii.gz
│   │   ├── Lumbar01_ClinicalCT_586X_586Y_1000Z_B_registered_BoneMask.nii.gz
│   │   └── Lumbar01_ClinicalCT_586X_586Y_1000Z_S_registered_BoneMask.nii.gz
│   └── ...
└── README.md

RegisteredData/ is the recommended starting point. BoneMask/ provides binary ROIs aligned to the released NIfTI volumes. Each sample contains one MicroPCCT mask and eight sequence-level clinical masks.

Data Statistics

Split Samples Clinical CT Micro-PCCT
Training / development 25 Available Available
Reference test set 5 Available Available

Input Sequences

For each sample, we provide 4 clinical CT sequences with soft tissue kernel:

Sequence Resolution Slice Thickness FOV
195X_195Y_500Z_S 195 μm 500 μm Small
195X_195Y_1000Z_S 195 μm 1000 μm Small
586X_586Y_500Z_S 586 μm 500 μm Large
586X_586Y_1000Z_S 586 μm 1000 μm Large

Micro-PCCT targets are released as int16 NIfTI volumes at 105 μm isotropic resolution.

Repository Contents

Directory Status Description
baseline/ Public Rigid registration baseline based on ANTs, including register_ants.py and batch_register.py
methods/ Public Super-resolution method scripts and usage notes
evaluation/ Public Image quality evaluation and trabecular summary scripts
docs/ Public Website pages, visualization assets, and benchmark result summaries

The public benchmark reports registered clinical CT, nearest interpolation, SRCNN, UNet, ESRGAN, and SwinIR under the released BoneMask-ROI protocol. Usage details are documented in methods/ and evaluation/.

Evaluation

CT Window Settings

All metrics are computed under three different CT window settings:

Window Window Center (WC) Window Width (WW) Purpose
Raw - - Original HU values (-1024 ~ 3071)
Bone 400 1800 Bone structure visualization
Soft Tissue 40 400 Soft tissue visualization

Metrics

Image-quality evaluation reports PSNR, SSIM, and MAE within the released BoneMask ROI.

Baseline Performance

Public benchmark tables cover registered clinical CT, Nearest, SRCNN, UNet, ESRGAN, and SwinIR.

Training uses Lumbar_01 to Lumbar_25; evaluation reports mean ± standard deviation on Lumbar_26 to Lumbar_30.

Quantitative comparison of super-resolution baselines across objective metrics and bone morphometry

Quantitative comparison of registration, image-quality, and trabecular morphometry results across the released baselines. Open the PDF version for the full-resolution figure.

View Full Baseline Results →

Registration Evaluation

Registration quality is reported within the released BoneMask ROI using Dice, HD95, and HD.

Method ROI Dice ↑ HD95 ↓ HD ↓
ANTs rigid baseline (195X_195Y_500Z_B) BoneMask ROI 0.9826 ± 0.0017 0.2266 ± 0.0131 mm 1.1476 ± 0.5960 mm
ANTs rigid baseline (586X_586Y_500Z_B) BoneMask ROI 0.9811 ± 0.0023 0.2887 ± 0.0358 mm 0.9095 ± 0.2725 mm
ANTs rigid baseline (overall) BoneMask ROI 0.9818 ± 0.0021 0.2576 ± 0.0411 mm 1.0285 ± 0.4785 mm

Bone Morphometry

The public benchmark reports ROI-based bone morphometry in the released BoneMask-defined VOI, together with paired comparison against Micro-PCCT.

Category Metrics
Core trabecular morphometry BV/TV, Tb.Th, Tb.Sp, Tb.N
Released VOI definition derived from the released BoneMask and used consistently for all public morphometry statistics

Current Public Subset: 195X_195Y_1000Z_S

Method / Data BV/TV Tb.Th (mm) Tb.Sp (mm) Tb.N (mm^-1)
Micro-PCCT reference 0.2208 ± 0.01740.2573 ± 0.01410.3647 ± 0.01920.8608 ± 0.0881
Registered clinical CT baseline 0.0055 ± 0.00290.5837 ± 0.11794.1319 ± 1.13140.0097 ± 0.0056
Δ / p-value vs Micro-PCCT -0.2125 / 0.03125+0.3761 / 0.03125+3.8776 / 0.03125-0.8398 / 0.03125
Nearest 0.0055 ± 0.00290.5837 ± 0.11794.1319 ± 1.13140.0097 ± 0.0056
Δ / p-value vs Micro-PCCT -0.2125 / 0.03125+0.3761 / 0.03125+3.8776 / 0.03125-0.8398 / 0.03125
SRCNN 0.0605 ± 0.02120.7673 ± 0.08650.9809 ± 0.24480.0773 ± 0.0202
Δ / p-value vs Micro-PCCT -0.1556 / 0.03125+0.5861 / 0.03125+0.5512 / 0.03125-0.7762 / 0.03125
UNet 0.0931 ± 0.02350.4324 ± 0.05500.4565 ± 0.05740.2133 ± 0.0346
Δ / p-value vs Micro-PCCT -0.1217 / 0.03125+0.1907 / 0.03125+0.0951 / 0.03125-0.6357 / 0.03125
ESRGAN 0.1330 ± 0.02390.2740 ± 0.01120.3941 ± 0.02490.4841 ± 0.0773
Δ / p-value vs Micro-PCCT -0.0878 / 0.03125+0.0167 / 0.03125+0.0293 / 0.09375-0.3767 / 0.03125
SwinIR 0.0415 ± 0.01400.7371 ± 0.08451.3594 ± 0.35290.0549 ± 0.0127
Δ / p-value vs Micro-PCCT -0.1793 / 0.03125+0.4798 / 0.03125+0.9947 / 0.03125-0.8059 / 0.03125

Current Public Subset: 586X_586Y_1000Z_S

Method / Data BV/TV Tb.Th (mm) Tb.Sp (mm) Tb.N (mm^-1)
Micro-PCCT reference 0.2208 ± 0.01740.2573 ± 0.01410.3647 ± 0.01920.8608 ± 0.0881
Registered clinical CT baseline 0.0026 ± 0.00180.5338 ± 0.13225.0502 ± 1.87190.0053 ± 0.0044
Δ / p-value vs Micro-PCCT -0.2167 / 0.03125+0.2665 / 0.03125+5.0664 / 0.03125-0.8493 / 0.03125
Nearest 0.0026 ± 0.00180.5338 ± 0.13225.0502 ± 1.87190.0053 ± 0.0044
Δ / p-value vs Micro-PCCT -0.2167 / 0.03125+0.2665 / 0.03125+5.0664 / 0.03125-0.8493 / 0.03125
SRCNN 0.1271 ± 0.02040.6565 ± 0.07390.4540 ± 0.04250.1928 ± 0.0104
Δ / p-value vs Micro-PCCT -0.0864 / 0.03125+0.4665 / 0.03125+0.0593 / 0.09375-0.6719 / 0.03125
UNet 0.1227 ± 0.02540.4639 ± 0.04670.4421 ± 0.04590.2624 ± 0.0300
Δ / p-value vs Micro-PCCT -0.0920 / 0.03125+0.2314 / 0.03125+0.0863 / 0.03125-0.5960 / 0.03125
ESRGAN 0.1506 ± 0.02090.2730 ± 0.00890.3710 ± 0.01600.5511 ± 0.0702
Δ / p-value vs Micro-PCCT -0.0702 / 0.03125+0.0156 / 0.03125+0.0063 / 0.40625-0.3097 / 0.03125
SwinIR 0.0486 ± 0.01440.4010 ± 0.02030.9216 ± 0.20530.1198 ± 0.0305
Δ / p-value vs Micro-PCCT -0.1722 / 0.03125+0.1437 / 0.03125+0.5568 / 0.03125-0.7410 / 0.03125

One-sided exact Wilcoxon signed-rank tests are reported against Micro-PCCT, paired by case and shown separately for each FOV (n = 5; smallest attainable exact one-sided p = 0.03125). Delta rows are reported as signed mean delta (pred - Micro-PCCT) / exact one-sided p value. Directions: BV/TV and Tb.N test pred < Micro-PCCT; Tb.Th and Tb.Sp test pred > Micro-PCCT.

Recommended Usage Policy

  1. Official Data: Use the released LumbarSR training set as the core development data.
  2. Disclosure: External data, pre-trained models, and synthetic data are allowed, but should be documented clearly.
  3. Integrity: Avoid hidden-ground-truth usage or any form of test leakage.

Authors and Affiliations

Ping Wang1,2, Ruipeng Zhang1,2, Mengfei Wang1,3, Shenyan Zong1,2, Jinyu Zhu1, Xinyu Song1, Zhenzhen Cao4, Xuefei Hu5, Dan Wang1*, Yuehua Li1,2*

  1. Institute of Diagnostic and Interventional Radiology, Shanghai Sixth People's Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, Shanghai, China
  2. Faculty of Medical Imaging Technology, College of Health Science and Technology, Shanghai Jiao Tong University School of Medicine
  3. School of Information and Intelligent Science, Donghua University, Shanghai, China
  4. Human Anatomy Lab, School of Basic Medical Science, Kunming Medical University, Kunming, China
  5. Tarim University School of Medicine, Alar, China

Ping Wang and Ruipeng Zhang contributed equally to this work.

Corresponding authors: Dan Wang; Yuehua Li

Contact

For questions about the dataset or benchmark, please contact the corresponding authors:
Dan Wang: joshuastonecn@sjtu.edu.cn
Yuehua Li: liyuehua77@sjtu.edu.cn
GitHub: @FrankZhangRp

Citation

If you use this dataset, please cite:

@misc{lumbarsr2026,
  title={LumbarSR: A Paired Clinical CT and Photon-Counting Micro-CT Dataset for Human Lumbar Vertebrae},
  author={Wang, Ping and Zhang, Ruipeng and Wang, Mengfei and Zong, Shenyan and Zhu, Jinyu and Song, Xinyu and Cao, Zhenzhen and Hu, Xuefei and Wang, Dan and Li, Yuehua},
  year={2026},
  url={https://github.com/FrankZhangRp/LumbarSR-Challenge}
}


LumbarSR: A Paired Clinical CT and Photon-Counting Micro-CT Dataset for Human Lumbar Vertebrae