Multimodal Imaging Dataset for Garden Classification of Femoral Neck Fractures

# Multimodal Imaging Dataset for Garden Classification of Femoral Neck Fractures

## Dataset Overview

This dataset is constructed for the intelligent Garden classification of femoral neck fractures, a core task for clinical precise diagnosis and treatment decision-making of hip fractures. It provides multi-modal hip CT imaging data covering all four Garden types (Type Ⅰ–Ⅳ) of femoral neck fractures, including both original CT images and implant occlusion-removed preprocessed images.

This dataset aims to fill the gap of publicly available, expert-annotated multi-modal orthopedic imaging datasets for artificial intelligence (AI)-driven fracture diagnosis research, and supports the development, training and validation of deep learning models for automated Garden classification of femoral neck fractures. This dataset is associated with the manuscript *Intelligent Garden Classification of Femoral Neck Fractures Using Multimodal Imaging Deep Learning Models for Clinical Precise Diagnosis*, which systematically evaluates the performance of 6 state-of-the-art deep learning models (StarNetS4, GhostNetV3, Swiftformer, FasterNet, SHViT, FastViT) on fracture classification tasks based on this multi-modal dataset.

 ## Ethics Statement

This retrospective study and the public release of this anonymized dataset were approved by the Medical Ethics Committee of Southwest University and the cooperating medical institutions (Beibei Hospital of Chongqing Medical University, the Ninth People's Hospital of Beibei Chongqing). All procedures were performed in full compliance with the ethical standards of the institutional research committee and the 1964 Helsinki Declaration and its later amendments. The requirement for written informed consent was waived by the ethics committee due to the retrospective, fully anonymized nature of the data, with all patient identifiable information completely removed to strictly protect patient privacy.

## Data Acquisition and Imaging Protocol

All imaging data were collected from cooperating medical institutions in Chongqing, China, from adult patients (age ≥ 18 years) with clinically and radiologically confirmed femoral neck fractures. The raw data are standardized spiral CT volumetric data with uniform scanning parameters:

- Tube voltage: 120 kV

- Tube current: 200–300 mA

 - Slice thickness/spacing: 0.625 mm

 - Matrix: 512×512

 - Field of view (FOV): 350×350 mm

Using the open-source medical image processing software 3D Slicer, three types of 2D images were reconstructed to simulate clinical routine reading perspectives, all included in this dataset:

1. **Axial CT views**: Cross-sectional slices centered on the femoral neck midpoint, clearly displaying cross-sectional fracture details, fracture line direction and fragment alignment;

2. **Coronal CT views**: Coronal plane slices of the femoral neck longest axis, corresponding to the core perspective of clinical anteroposterior and lateral X-rays;

3. **3D projection views**: Maximum Intensity Projection (MIP) images with a 30° projection angle relative to the coronal plane and 10 mm projection thickness, displaying the overall 3D morphological characteristics of fractures.

 ## Dataset Composition

The dataset is split into training, validation, and independent test sets at a ratio of 7:1.5:1.5 at the patient level, to avoid data leakage in model training. All images and data from the same patient are assigned to only one subset, with balanced baseline characteristics (gender, age, fracture type distribution) across all subsets (P > 0.05). The detailed sample distribution is as follows:

 | Imaging Modality | Training Set | Validation Set | Test Set | Garden Type Distribution (Ⅰ/Ⅱ/Ⅲ/Ⅳ) | |-------------------|--------------|----------------|----------|---------------------------------------|

| Original Axial | 2547 | 652 | 623 | ~15.4% / ~26.0% / ~28.2% / ~30.4% |

| Original Coronal | 2601 | 640 | 1207 | ~15.5% / ~25.8% / ~28.5% / ~30.2% |

| Original 3D Projection | 1108 | 271 | 471 | ~15.3% / ~26.1% / ~28.7% / ~29.9% |

| Occlusion-Removed Axial | 2543 | 596 | 611 | ~15.4% / ~25.9% / ~28.3% / ~30.4% |

| Occlusion-Removed Coronal | 2460 | 613 | 1147 | ~15.6% / ~25.9% / ~28.4% / ~30.1% |

| Occlusion-Removed 3D Projection | 1108 | 271 | 471 | ~15.3% / ~26.1% / ~28.7% / ~29.9% |

The dataset also includes the expert-annotated gold standard of Garden classification for all samples, as well as the complete preprocessing code for image normalization and implant occlusion removal.

 ## Annotation Standard and Quality Control

All Garden classification annotations were completed by two associate chief orthopedic surgeons with over 10 years of clinical experience, strictly following the standard Garden classification system for femoral neck fractures. Annotators performed blinded independent reading without access to the patient's clinical diagnosis and treatment plan.

Discrepancies with classification differences ≥ 1 type were resolved by a third chief orthopedic surgeon with over 20 years of experience, who determined the final gold standard classification by combining multimodal images and complete clinical information. Inter-annotator consistency was evaluated by Cohen's Kappa coefficient: the Kappa coefficient between the two primary annotators was 0.82 (95% CI: 0.78–0.86), and the Kappa coefficient between the final annotation and the gold standard reached 0.95 (95% CI: 0.93–0.97), indicating excellent and reliable annotation quality.

Preprocessing Pipeline

All images in the dataset were processed with a standardized preprocessing pipeline:

1. All reconstructed 2D images were uniformly resized to 224×224 pixels;

2. Pixel values were normalized to [0,1] using Z-score standardization to eliminate systematic differences caused by scanning parameters;

3. For images with internal fixator occlusion, a dedicated preprocessing pipeline was applied: a U-Net++ model was used for high-precision segmentation of occluded regions, followed by a PatchGAN inpainting algorithm to restore anatomical structure and fracture line details, with a segmentation accuracy of 96.8% for occluded regions.

## Citation

When using this dataset, please cite the associated manuscript: Wu, M., Ma, Y., Li, Y., Zhang, X., Zhu, Y., Zhou, J., & Chen, S. (2026). Intelligent Garden Classification of Femoral Neck Fractures Using Multimodal Imaging Deep Learning Models for Clinical Precise Diagnosis.

## License

This dataset is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0). Users are free to share and adapt the dataset, provided that appropriate credit is given to the original authors, a link to the license is provided, and any changes made are clearly indicated.