FracAtlas Dataset for Fracture Detection

• FracAtlas Dataset is a curated collection of 4,083 musculoskeletal X-rays from leg, hand, and hip with significant class imbalance between fractured and non-fractured cases.
• The dataset undergoes standardized preprocessing including resizing, normalization, and augmentation to align with both custom CNN and transfer learning model requirements.
• Custom CNN models trained on FracAtlas achieve high accuracy (95.96%), outperforming transfer learning approaches in tackling fracture detection biases.

FracAtlas Dataset is a publicly available collection of medical radiographs specifically curated for the development, benchmarking, and validation of AI-based fracture detection algorithms. The dataset consists of 4,083 anonymized musculoskeletal X-ray images obtained from three distinct anatomical regions: leg, hand, and hip. It is particularly noted for its pronounced class imbalance, with 3,366 non-fractured and 717 fractured cases. FracAtlas is employed as a primary resource in the design, training, and comparative analysis of custom and transfer learning convolutional neural network (CNN) models for automated bone fracture detection. Its composition, curation, and application in AI research reflect both its utility and its limitations for translating machine learning prototypes into clinical practice.

1. Dataset Composition and Structure

FracAtlas comprises 4,083 de-identified X-ray images categorized into fractured and non-fractured classes. The images originate from three anatomical regions:

• Leg
• Hand
• Hip

A key characteristic is the significant class imbalance—only 717 of the cases are fractured, with the remainder (3,366) being non-fractured. All images are provided in resolutions appropriate for deep learning workflows. For model compatibility, images were standardized to 128×128 pixels for custom CNNs and 224×224 pixels for transfer learning models.

Anatomical Region	Fractured Cases	Non-Fractured Cases
Leg	—	—
Hand	—	—
Hip	—	—
Total	717	3,366

(Exact anatomical breakdown not specified; the table reflects structure as documented.)

This imbalance necessitates specific training strategies to prevent model bias towards the majority class.

2. Image Preprocessing and Augmentation

All images underwent a standardized preprocessing pipeline prior to AI model ingestion:

• Resizing: To match model input dimensions (128×128 for custom CNN, 224×224 for transfer learning architectures).
• Conversion: Images were converted to arrays using common utilities such as img_to_array.
• Normalization: Pixel values were normalized by dividing by 255.

To address the class imbalance, augmentation techniques including rotation, zooming, and flipping were applied. Stratified sampling was used to split the dataset into training and validation sets, preserving the proportion of fractured to non-fractured instances in both splits. This strategy counteracts the skewed distribution and improves the generalizability of models trained on FracAtlas.

3. Integration in AI-Based Fracture Detection Systems

FracAtlas serves as both the training and evaluation dataset in AI-based fracture detection studies. The typical workflow involves:

• Developing a custom CNN or leveraging transfer learning models (EfficientNetB0, MobileNetV2, ResNet50).
• Fine-tuning model input pipelines to the FracAtlas data structure.
• Applying stratified data splits and augmentation to preserve clinical relevance and mitigate overfitting.
• Benchmarking models using metrics such as accuracy, precision, recall, and F1-score.

For example, a custom sequential Keras CNN comprising three convolutional blocks (Conv2D, batch normalization, max pooling, dropout), a flattening stage, dense layers, and a single neuron with sigmoid output is specifically optimized with the Adam optimizer and binary cross-entropy loss function:

$$L = -\frac{1}{N}\sum_{i=1}^{N}\left[y_i \log(\hat{y}_i) + (1-y_i)\log\left(1-\hat{y}_i\right)\right]$$

Transfer learning models are adapted by replacing final classification layers with custom dense modules and freezing base model weights learned from ImageNet.

4. Benchmark Results and Model Performance

Studies utilizing the FracAtlas dataset report the following key performance metrics for their custom CNN models:

• Accuracy: 95.96%
• Precision: 0.94
• Recall: 0.88
• F1-score: 0.91

Transfer learning models—EfficientNetB0, MobileNetV2, ResNet50—demonstrate poor performance under the highly imbalanced FracAtlas setup, with fractured class precision values as low as 0.23–0.27 and recall between 0.38 and 0.53. This inferior performance is attributed to both the dataset’s size and the skewed class distribution.

The high performance of custom, lightweight CNNs trained specifically on FracAtlas indicates that domain-tuned architectures may be more effective for imbalanced and domain-constrained medical imaging tasks compared to broadly pre-trained deep models.

5. Limitations and Implications for Clinical Translation

FracAtlas’s composition imposes several notable limitations:

• Limited Diversity: Only three anatomical regions (leg, hand, hip) are included, restricting the generalizability of learned models.
• Class Imbalance: The significant skew impacts transfer learning approaches and can bias results.
• No External Validation: All reported results pertain to internal validation splits; clinically credible AI systems require external, multi-center validation on more heterogeneous data.
• Scale: Dataset size is modest relative to contemporary medical imaging repositories, further impacting model robustness and generalization.

A plausible implication is that while FracAtlas supports rapid prototyping and benchmarking for fracture detection, it should ideally be supplemented or followed by more comprehensive, diverse, and balanced datasets to support clinical-grade model deployment.

6. Recommendations for Future Research

Documented recommendations include:

• Expand anatomical coverage to additional skeletal regions and diverse fracture typologies.
• Increase dataset size via new image acquisitions to ensure better model generalizability.
• Support multimodal imaging (CT, MRI) to enable multi-class and multi-modality clinical tasks.
• Prioritize external validation through multi-center studies and broader data sharing, thereby enhancing clinical translation.
• Adopt more sophisticated sampling and augmentation to counteract class imbalance.

Addressing these points will enable FracAtlas to serve as a foundation for more robust, generalizable, and clinically relevant AI-based fracture detection systems and facilitate ongoing innovation in both model development and deployment.