BoneFract: A Bone Fracture Dataset
Description
The research hypothesis underlying the BoneFract dataset is that a large, patient-level collection of labeled bone X-ray images can provide a useful resource for developing and evaluating machine learning and deep learning methods for bone fracture detection and anatomical-region classification. The dataset contains 47,931 X-ray images from 47,931 unique patients, with one original X-ray image per patient. The images were collected from the Peoples University of Medical & Health Sciences for Women (PUMHSW), Nawabshah, Sindh, Pakistan. The dataset covers seven anatomical regions: Arm, Wrist, Hand, Pelvis, Thigh, Lower Leg, and Foot. Each image is labeled as Positive or Negative according to fracture status. The dataset contains 13,921 Positive and 34,010 Negative images. The largest anatomical category is Lower Leg, while the distribution of fracture-positive and fracture-negative images varies across anatomical regions. This class distribution should be considered when designing experiments and interpreting model performance, particularly for class-wise evaluation. The images are provided in PNG format and standardized to 512 × 512 pixels. The dataset contains natural variations in illumination, exposure, contrast, patient positioning, imaging systems, and acquisition conditions. These variations provide diversity in image appearance and can be used to evaluate the performance and robustness of automated image-classification methods. The dataset is divided into predefined training, validation, and test sets, containing 38,344, 4,793, and 4,794 images, respectively. The training set can be used for model development, the validation set for model selection and parameter tuning, and the test set for independent performance evaluation. The predefined partitions should be maintained when comparing experimental results. A Positive label represents an X-ray image categorized as showing a fracture, whereas a Negative label represents an image categorized as non-fractured within the respective anatomical region. The anatomical class identifies the region represented in the X-ray image. Researchers can therefore use the dataset for binary fracture-status classification, anatomical-region classification, or multiclass classification tasks. All released images are original selected X-ray records, and no augmented images are included in the dataset. Patient-identifying information was removed before release. The dataset can be used for research in medical image analysis, computer vision, machine learning, transfer learning, and deep learning, including model development, benchmarking, class-wise performance analysis, and investigation of automated bone fracture classification. The dataset should be interpreted as a research dataset and not as a clinically validated diagnostic tool. Model predictions obtained from the dataset should not be used as a substitute for professional medical assessment.
Files
Institutions
- People's University of Medical and Health Sciences for WomenSindh, Nawabshah