Endplate3D-QCT: A High-Resolution Dataset

and Benchmark for Automated 3D Segmentation

of Lumbar Vertebral Endplates in QCT

Zixun Yin

, Da Zou

, Yi Zhao

, Chenbin Zhang

, Weishi Li

, Minghui Wu

1,6

Kun Yan

3,4*

, and Ping Wang

1,4,5*

School of Software and Microelectronics, Peking University, Beijing, China

Department of Orthopaedics, Peking University Third Hospital, Beijing, China

School of Computer Science, Peking University, Beijing, China

National Engineering Research Center for Software Engineering, Peking University,

Beijing, China

Key Laboratory of High Conﬁdence Software Technologies (PKU), Ministry of

Education, China

Shanghai Artiﬁcial Intelligence Laboratory and Mininglamp Technology, China

Abstract. Accurate segmentation of lumbar vertebral endplates is es-

sential for assessing bone density and biomechanical properties in spinal

disorders. While quantitative computed tomography (QCT) provides de-

tailed bone density measurements, existing segmentation approaches pri-

marily focus on vertebral bodies and intervertebral discs, often neglect-

ing the precise delineation of endplates. Current deep learning methods

perform well in healthy spines but struggle with pathological cases due

to the thin and morphologically complex nature of endplates, particu-

larly in the presence of osteophytes and degenerative changes. To ad-

dress these challenges, we introduce the ﬁrst publicly available dataset,

Endplate3D-QCT, which contains pixel-level annotations of lumbar end-

plates in clinical QCT scans. Our dataset includes high-precision 3D

segmentation masks targeting cortical endplates and subchondral bone,

along with an automated evaluation framework for model assessment.

We benchmark multiple deep learning models, including EﬃcientUNet,

UNet, VNet, UNETR and SwinUNETR, using nnUNet as the train-

ing framework. While these models achieve Dice scores around 0.9, they

exhibit inconsistencies in endplate identiﬁcation, leading to false posi-

tives and false negatives. These ﬁndings highlight the need for further

advancements in endplate segmentation techniques. Our dataset and

benchmarks provide a valuable foundation for improving spinal implant

design, bone density mapping, and computational modeling of vertebral

load distribution. The dataset and the evaluation code are available at

https://github.com/yin876705249/Endplate3D-QCT.

Keywords: Lumbar Endplate Segmentation · Open-access Dataset ·

Evaluation Framework

Corresponding authors: Kun Yan (kyan2018@pku.edu.cn) and Ping Wang

(pwang@pku.edu.cn).

2 Z. Yin and D. Zou et al.

1 Introduction

Accurate segmentation of lumbar vertebral endplates is critical for quantitative

assessment of bone density and biomechanical characterization in spinal dis-

orders. While quantitative computed tomography (QCT) enables density mea-

surements of trabecular and cortical bone, existing segmentation methods for

spinal structures—typically focusing on intervertebral discs and vertebral bod-

ies—fail to achieve the precision required for endplate-speciﬁc analysis. Current

deep learning approaches often yield satisfactory results in healthy spines with

minimal osteophytes but struggle to delineate the thin (<4 mm), anatomically

ambiguous endplate regions in pathological cases. This limitation stems from two

key factors: 1) the reliance on threshold-based 2D methods that inadequately

capture 3D morphological variations of the endplate’s cortical bone layer, and

2) the lack of high-resolution CT datasets with expert annotations speciﬁcally

targeting the osseous endplate-cartilage interface.

The clinical imperative of endplate segmentation lies in its direct relevance

to spinal surgery outcomes. During procedures like interbody fusion, where im-

plants contact the endplate surface, local bone density measurements could guide

personalized device selection and 3D-printed cage design—critical for preventing

subsidence in osteoporotic patients. Furthermore, precise endplate characteriza-

tion could advance research on asymmetric bone remodeling in scoliosis and

degenerative disc disease. While MRI excels at visualizing soft tissues like discs

and neural structures [3], CT remains the modality of choice for osseous endplate

analysis due to superior spatial resolution and contrast for cortical bone.

Despite these applications, no publicly available dataset currently provides

detailed 3D annotations of lumbar endplates. Existing spinal CT segmentation

works [12, 10] predominantly focus on vertebral bodies and discs, often treating

the endplate as a byproduct of post-processing rather than a distinct anatomical

entity. This gap hinders the development of machine learning models capable of

capturing the endplate’s subtle radiological signatures, particularly in degener-

ative spines where pathological changes obscure traditional intensity-based seg-

mentation boundaries.

To address these challenges, we present the ﬁrst open-access dataset with

pixel-level annotations of lumbar endplates in clinical QCT scans. Our contri-

butions include: 1) High-precision 3D segmentation masks delineating cortical

endplates and subchondral bone regions. 2) An automated evaluation framework

for precise assessment of segmentation model performance. 3) Comprehensive

benchmarking experiments comparing widely-used segmentation methods.

This resource aims to accelerate research in personalized spinal implant de-

sign, bone density mapping, and computational modeling of load distribution

across the vertebral column. We conducted benchmark experiments based on

nnUNet [6, 7], a state-of-the-art framework in the medical image segmentation

domain. We evaluated the performance of widely used segmentation methods,

including UNet [11], VNet [9], and SwinUNETR [4], etc. The results indicate

that while these methods achieve high segmentation performance (with Dice

scores around 0.9), they do not consistently ensure the accurate identiﬁcation

Endplate3D-QCT: A High-Resolution Dataset and Benchmark 3

of all endplates, leading to instances of false negatives and false positives. This

suggests that endplate segmentation remains a challenging task that requires

further exploration.

2 Dataset Construction

2.1 Data Acquisition

Our study utilized 119 CT volumes from the CTSpine1K dataset [2] containing

complete lumbar regions, which were carefully selected based on image qual-

ity and anatomical completeness. After extracting volumetric regions of interest

(ROIs) centered on the lumbar spine, we performed pixel-level 3D annotations

of vertebral endplates following clinical guidelines, including superior/inferior

endplates from L1 to L5 and the superior endplate of S1. This resulted in 1,309

annotated endplate surfaces (11 per case) with sub-millimeter precision, ad-

dressing the critical gap in publicly available endplate segmentation data. To

our knowledge, this represents the ﬁrst open-access dataset providing large-scale

(over 1,000 annotated surfaces), high-quality 3D annotations of lumbar end-

plates, particularly valuable for developing AI-assisted surgical planning systems

and biomechanical modeling in spinal disorders.

In addition to the re-annotated CTSpine-Reﬁned dataset, we collected an-

other 51 patients’ preoperative lumbar CT images in this study. These patients

were clinically diagnosed as lumbar degenerative diseases and underwent lumbar

fusion surgery. The scanning parameters of CT were set at the values of 120 kVp.

The patients group included 23 males and 28 females, and their average age was

66.2 ± 5.3 years, ranging from 51 to 76 years. Concretely, we constructed two

datasets from these 51 clinical cases (10 with scoliosis) of lumbar degenerative

diseases (LDD) reﬂecting diﬀerent levels of segmentation diﬃculty:

1. LDD-Mild includes 25 cases with mild symptoms and regular endplates,

representing a relatively low level of segmentation complexity.

2. LDD-Severe includes 26 cases with severe symptoms and irregular endplates,

posing signiﬁcant challenges for segmentation models. These datasets serve

to test model robustness across a spectrum of anatomical variations.

In this study, the cortical bone of upper and lower lumbar endplates from L1

to L5 together with S1 upper endplates were segmented on lumbar CT images.

The segmentation process was performed in the sagittal plane of CT images slice

by slice to annotate the whole endplate, as shown in Fig. 1. Five spine surgeons

participated in the CT image annotation. Four surgeons with over 5 years of

clinical experience performed the initial segmentation. All received standardized

training and strictly followed the annotation manual to minimize bias. Using

cross-annotation and iterative feedback, the team held multiple discussions to

resolve inconsistencies. Additionally, a spine surgeon with more than 10 years of

experience reexamined all annotated images, further ensuring quality. Detailed

information of our Endplate3D-QCT can be found in Table 1. Our dataset con-

sists of a total of 170 scans, with each scan containing instance segmentation

4 Z. Yin and D. Zou et al.

annotations for 11 endplates. In other words, we provide a total of 1,870 an-

notated endplate instances, making this the largest publicly available endplate

segmentation dataset to date.

Table 1. Summary statistics of the Endplate3D-QCT dataset. BMD (mg/cm

): Bone

Mineral Density measured in endplate region.

Metric CTSpine-Reﬁned LDD-Mild LDD-Severe Overall

Image Count 119 25 26 170

Avg Image Size (190, 188, 295) (204, 209, 305) (199, 202, 295) -

Spacing (mm) (0.8, 0.8, 0.8) (0.4, 0.4, 0.75) (0.39 0.39 0.74) -

L1 endplate count 238 50 52 340

L1 endplate BMD 129.8 ± 6.8 128.8 ± 7.9 130.1 ± 7.0 129.7 ± 7.0

L2 endplate Count 238 50 52 340

L2 endplate BMD 129.0 ± 6.4 127.8 ± 7.2 129.9 ± 6.5 129.0 ± 6.6

L3 endplate Count 238 50 52 340

L3 endplate BMD 129.0 ± 5.9 127.6 ± 6.8 129.2 ± 5.7 128.8 ± 6.1

L4 endplate Count 238 50 52 340

L4 endplate BMD 128.9 ± 5.3 127.7 ± 5.8 128.4 ± 4.9 128.6 ± 5.3

L5 endplate Count 238 50 52 340

L5 endplate BMD 128.5 ± 6.1 128.3 ± 5.0 129.0 ± 4.3 128.6 ± 5.7

S1 endplate Count 119 25 26 170

S1 endplate BMD 128.4 ± 6.2 126.8 ± 4.7 127.8 ± 4.7 128.0 ± 7.3

2.2 Evaluation Protocol

Our automated evaluation framework employs multi-instance morphological anal-

ysis with clinical interpretability, formalized through the following mathematical

constructs.

Volumetric Metrics Given a predicted segmentation mask P and ground truth

G in 3D space Ω ⊂ Z

, we deﬁne:

DSC(P, G) =

2|P ∩ G|

|P | + |G|

∈ [0, 1], JI(P, G) =

|P ∩ G|

|P ∪ G|

∈ [0, 1] (1)

For surface distance metrics, let S

and S

denote the lumber endplate

surfaces with point sets {p

} and {g

}, respectively:

(P, G) = max



95th

p∈S

d(p, S

), 95th

g∈S

d(g, S

)



(2)

Endplate3D-QCT: A High-Resolution Dataset and Benchmark 5

Fig. 1. Visualization of Endplate3D-QCT. The images in CTSpine-Reﬁned were ac-

quired from healthy individuals, with the alignment between vertebral bodies better

conforming to the natural curvature of the human spine. In contrast, LDD-Mild exhibits

a certain degree of degeneration, with some vertebrae showing signs of deformation.

LDD-Severe represents more advanced degeneration, where some endplates have even

undergone deterioration.

ASD(P, G) =

| + |S





i=1

d(p

, S

) +

j=1

d(g

, S

)





(3)

Instance Matching For multi-instance endplate detection, deﬁne the optimal

assignment matrix A ∈ {0, 1}

M×N

where M predicted instances and N ground

truth instances:

(

1, if arg max

DSC(P

, G

) = j and DSC(P

, G

) > 0.8

0, otherwise

(4)

Precision, recall and F1 are then computed as:

Precision =

i=1

j=1

, Recall =

j=1

i=1

(5)

and F1 is obtained through: F1 = 2 × (Precision × Recall)/(Precision + Recall).

3 Experiments

3.1 Implementation Details

We conducted experiments based on nnUNet, state-of-the-art training frame-

work, and evaluate ﬁve widely-used methods, including UNet [11], VNet [9],

6 Z. Yin and D. Zou et al.

Fig. 2. Visualization of segmentation results. The ﬁrst row presents a sample from

CTSpine-Reﬁned, where the models generally exhibit missed detections on the S1 end-

plate. The second row corresponds to LDD-Mild; due to the higher data resolution,

the models generally achieve better performance. The third row shows examples from

LDD-Severe, where severe endplate degeneration leads to poor model performance,

particularly for EﬃcientNet and UNet (2D).

UNETR [5], SwinUNETR [4] and EﬃcientUNet [1]. All models were trained us-

ing Adam optimizer [8] with initial learning rate 3e-4 and cosine decay schedule.

We implemented a hybrid loss function combining dice loss and cross entropy

loss. For 3D methods, we used patch-based training and test-time sliding window

inference. Data augmentation included random cropping, rotation, and scaling.

Detailed implementation of each segmentation methods is shown in Table. 2. We

performed three-fold cross-validation on the CTSpine-Reﬁned subset and used

an ensemble of the three models to make inferences on the other two subsets.

When evaluating each subset, we ﬁrst calculate the instance-level Precision,

Recall, and F1-score for endplates using Eq.5, which we refer to as the "Detec-

tion" metrics. Second, we compute the overall segmentation performance using

Eq.1, which we deﬁne as the "Segmentation" metrics. Finally, we assess the seg-

mentation performance for each individual endplate instance, including Dice,

Jaccard, HD95, and ASD, which we categorize as the "Instance Segmentation"

metrics.

3.2 Dataset-Centric Performance Analysis

Our experimental evaluation focuses on revealing how dataset characteristics in-

ﬂuence segmentation performance across diﬀerent anatomical presentations. We

Endplate3D-QCT: A High-Resolution Dataset and Benchmark 7

Table 2. Implementation details of each method.

Model Parameters (M) Input Size FLOPs (GFLOPs)

EﬃcientUNet [1] 10.05 (288, 288) 3.20

UNet (2D) [11] 18.67 (288, 288) 19.78

UNETR [5] 121.51 (320, 96, 96) 213.28

SwinUNETR [4] 15.7 (224, 96, 96) 184.01

UNet (3D) [11] 31.19 (320, 96, 96) 749.42

VNet [9] 9.46 (320, 96, 96) 178.85

Table 3. Performance comparison on CTSpine-Reﬁned subset of our Endplate3D-

QCT. The units of HD95 and ASD are millimeters (mm).

Method

Detection Segmentation Instance Segmentation

Precision Recall F1 Dice Jaccard Dice Jaccard HD95 ASD

EﬃcientUNet 0.904 0.909 0.906 0.867 0.766 0.878 0.785 1.007 0.258

UNet(2D) 0.945 0.942 0.943 0.885 0.795 0.891 0.804 1.003 0.225

UNETR 0.923 0.926 0.925 0.878 0.784 0.888 0.800 1.012 0.245

SwinUNETR 0.923 0.945 0.934 0.890 0.803 0.904 0.827 1.004 0.195

UNet(3D) 0.948 0.946 0.947 0.901 0.821 0.905 0.827 1.002 0.202

VNet 0.946 0.949 0.947 0.901 0.821 0.905 0.828 1.003 0.201

analyze three clinically distinct subsets: CTSpine-Reﬁned (healthy morphology),

LDD-Mild (early degeneration), and LDD-Severe (advanced pathology).

CTSpine-Reﬁned: Baseline Performance. In cases with preserved anatomy

(0.8mm isotropic voxels), all 3D methods achieved ≥ 0.92 F1 score for endplate

detection. The uniform intensity proﬁles (HU range 1,200-1,500) enabled reliable

boundary identiﬁcation, with surface distance metrics showing submillimeter

accuracy (ASD = 0.195–0.258mm).

However, even in this optimal scenario, as shown in Fig. 2, we observed false

negatives in S1 endplate detection due to sacral curvature variations, highlighting

inherent anatomical complexity.

LDD-Mild: Degenerative Onset. The 0.4mm axial resolution in early de-

generation cases improved vertical structure visualization. Overall, the perfor-

mance of diﬀerent methods on the LDD-Mild subset is slightly better than on

the CTSpine-Reﬁned subset. This is mainly due to the higher resolution of im-

ages in LDD-Mild, which provides more detailed information. Additionally, the

endplate degeneration in LDD-Mild is generally mild, posing less of a challenge

for the segmentation models.

8 Z. Yin and D. Zou et al.

Table 4. Performance comparison on LDD-Mild subset of our Endplate3D-QCT. The

units of HD95 and ASD are millimeters (mm).

Method

Detection Segmentation Instance Segmentation

Precision Recall F1 Dice Jaccard Dice Jaccard HD95 ASD

EﬃcientUNet 0.851 0.868 0.860 0.872 0.775 0.884 0.794 1.035 0.219

UNet (2D) 0.915 0.938 0.926 0.887 0.797 0.891 0.805 1.008 0.225

UNETR 0.947 0.947 0.947 0.895 0.810 0.898 0.816 1.018 0.233

SwinUNETR 0.946 0.949 0.947 0.905 0.827 0.909 0.835 1.014 0.201

UNet (3D) 0.972 0.961 0.966 0.915 0.844 0.916 0.847 1.005 0.187

VNet 0.960 0.956 0.958 0.918 0.849 0.919 0.852 1.010 0.187

Table 5. Performance comparison on LDD-Severe subset of our Endplate3D-QCT.

The units of HD95 and ASD are millimeters (mm).

Method

Detection Segmentation Instance Segmentation

Precision Recall F1 Dice Jaccard Dice Jaccard HD95 ASD

EﬃcientUNet 0.600 0.554 0.576 0.773 0.641 0.862 0.759 1.123 0.232

UNet (2D) 0.704 0.728 0.716 0.832 0.716 0.869 0.770 1.139 0.273

UNETR 0.778 0.785 0.782 0.853 0.747 0.880 0.788 1.139 0.294

SwinUNETR 0.803 0.818 0.811 0.863 0.761 0.886 0.798 1.120 0.261

UNet (3D) 0.887 0.891 0.889 0.880 0.787 0.895 0.813 1.112 0.258

VNet 0.875 0.870 0.872 0.885 0.796 0.900 0.821 1.125 0.253

LDD-Severe: Pathological Extremes. Severely degenerated cases (0.39mm

spacing, extensive osteophytes) exposed fundamental limitations. Despite having

the same voxel dimensions as LDD-Mild, the increased pathological complexity

led to a decline in all evaluation metrics, highlighting the challenges of LDD-

Severe. Although the Segmentation performance did not drop signiﬁcantly (e.g.,

Dice decreased from 0.918 to 0.885 for VNet), the decline in Detection metrics

was particularly pronounced, with Recall dropping below 0.9 across all methods.

Notably, EﬃcientUNet had the lowest Recall, reaching only 0.554. We believe

that the decline in performance of all models on this subset is primarily due to

the excessive morphological variability of the endplates in these cases, resulting

in their irregularity.

4 Conclusion

In this study, we introduce the open-access dataset Endplate3D-QCT, featuring

pixel-level annotations of lumbar endplates in clinical quantitative CT (QCT)

scans. Building upon this dataset, we developed an automated evaluation frame-

work for robust performance assessment. Extensive experiments were conducted

using widely-used segmentation models to perform comprehensive benchmark-

Endplate3D-QCT: A High-Resolution Dataset and Benchmark 9

ing. These resources establish a foundation for advancing endplate-speciﬁc anal-

ysis, contributing to improved research in spinal biomechanics, personalized im-

plant design, and bone density mapping.

While our dataset represents a signiﬁcant advancement in endplate charac-

terization, current deep learning methods still fall short of the precision required

for clinical application. Despite achieving high Dice scores on relatively healthy

spines, all tested models exhibit inconsistencies in identifying endplates, partic-

ularly in the presence of severe degenerative changes. To bridge the gap between

research and clinical applicability, future eﬀorts should develop more robust seg-

mentation techniques capable of handling pathological variations. Furthermore,

intraoperative validation studies are essential to establish direct correlations be-

tween segmentation accuracy and surgical outcomes, ensuring that AI-assisted

methods meet the stringent requirements of clinical decision-making.

Acknowledgments. This research was supported in part by the Brain-like General

Vision Model and Applications project (2022ZD0160403).

Disclosure of Interests. The authors have no competing interests to declare that

are relevant to the content of this paper.

References

1. Baheti, B., Innani, S., Gajre, S., Talbar, S.: Eﬀ-unet: A novel architecture

for semantic segmentation in unstructured environment. In: Proceedings of the

IEEE/CVF conference on computer vision and pattern recognition workshops. pp.

358–359 (2020)

2. Deng, Y., Wang, C., Hui, Y., Li, Q., Li, J., Luo, S., Sun, M., Quan, Q., Yang, S.,

Hao, Y., et al.: Ctspine1k: A large-scale dataset for spinal vertebrae segmentation

in computed tomography. arXiv preprint arXiv:2105.14711 (2021)

3. van der Graaf, J.W., van Hooﬀ, M.L., Buckens, C.F., Rutten, M., van Susante,

J.L., Kroeze, R.J., de Kleuver, M., van Ginneken, B., Lessmann, N.: Lumbar spine

segmentation in mr images: a dataset and a public benchmark. Scientiﬁc Data

11(1), 264 (2024)

4. Hatamizadeh, A., Nath, V., Tang, Y., Yang, D., Roth, H., Xu, D.: Swin unetr:

Swin transformers for semantic segmentation of brain tumors in mri images (2022),

https://arxiv.org/abs/2201.01266

5. Hatamizadeh, A., Tang, Y., Nath, V., Yang, D., Myronenko, A., Landman, B.,

Roth, H., Xu, D.: Unetr: Transformers for 3d medical image segmentation (2021),

https://arxiv.org/abs/2103.10504

6. Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a

self-conﬁguring method for deep learning-based biomedical image segmentation.

Nature methods 18(2), 203–211 (2021)

7. Juanying, X., Ying, P., Mingzhao, W.: The squeeze & excitation normal-

ization based nnu-net for segmenting head & neck tumors. Chinese Journal

of Electronics 33(3), 766–775 (2024). https://doi.org/10.23919/cje.2022.00.306,

https://cje.ejournal.org.cn/en/article/doi/10.23919/cje.2022.00.306

8. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint

arXiv:1412.6980 (2014)

10 Z. Yin and D. Zou et al.

9. Milletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural net-

works for volumetric medical image segmentation. arXiv preprint arXiv:1606.04797

(2016)

10. Möller, H., Graf, R., Schmitt, J., Keinert, B., Schön, H., Atad, M., Sekuboyina, A.,

Streckenbach, F., Koﬂer, F., Kroencke, T., et al.: Spineps—automatic whole spine

segmentation of t2-weighted mr images using a two-phase approach to multi-class

semantic and instance segmentation. European Radiology pp. 1–12 (2024)

11. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed-

ical image segmentation. arXiv preprint arXiv:1505.04597 (2015)

12. Zhang, Z., Liu, T., Fan, G., Li, B., Feng, Q., Zhou, S.: Spinemamba: Enhancing 3d

spinal segmentation in clinical imaging through residual visual mamba layers and

shape priors. arXiv preprint arXiv:2408.15887 (2024)