Multi-Class Knee Osteoporosis X-ray Dataset

This dataset contains knee radiographic (X-ray) images collected and reorganized for osteoporosis classification research using deep learning techniques. The primary dataset was obtained from Kaggle (https://www.kaggle.com/datasets/866059b7930a5c49cd77d94c1761840a19d88074cad74e8f0e0cfa2b236a6904) and consists of three classes representing bone density conditions: Normal, Osteopenia, and Osteoporosis. The dataset contains a total of 1,947 knee X-ray images, including 780 Normal images, 793 Osteoporosis images, and 374 Osteopenia images, making it suitable for osteoporosis classification and medical image analysis tasks.In addition, the Knee X-ray Osteoporosis Database dataset was also used and this dataset was collected by Insha Majeed Wani and Sakshi Arora from Shri Mata Vaishno Devi University, this dataset was also published on Mendeley Data https://data.mendeley.com/datasets/fxjm8fb6mw/2. The Knee X-ray Osteoporosis Database contains 350 knee X-ray images divided into three classes: Normal, Osteoporosis, and Osteopenia. This dataset is imbalanced in class distribution with only 36 normal images, 154 osteopeni,a and 49 osteoporosis.

Since both datasets exhibited class imbalance, several data augmentation techniques were applied to improve class distribution and increase data diversity. These augmentation methods included random horizontal and vertical flipping and random image rotations at different angles. Augmentation was performed to generate balanced image samples across all classes and improve the robustness of the deep learning model. After preprocessing and augmentation, each class was balanced to contain an equal number of images.

The final combined dataset was divided into training, validation, and testing subsets using a ratio of 70%, 20%, and 10%, respectively. The prepared dataset was then used for osteoporosis classification using deep learning techniques on knee radiographic images.

本数据集为基于深度学习技术开展骨质疏松症分类研究而收集整理的膝关节放射（X-ray）影像数据集。其核心数据集源自Kaggle平台（链接：https://www.kaggle.com/datasets/866059b7930a5c49cd77d94c1761840a19d88074cad74e8f0e0cfa2b236a6904），包含三类骨密度状态类别：正常（Normal）、骨量减少（Osteopenia）以及骨质疏松（Osteoporosis）。该数据集总计包含1947张膝关节X线影像，其中正常影像780张、骨质疏松影像793张、骨量减少影像374张，适用于骨质疏松症分类及医学影像分析相关任务。
此外，本研究还使用了由印莎·马吉德·瓦尼（Insha Majeed Wani）与萨克希·阿罗拉（Sakshi Arora）从什里·马塔·瓦伊什诺·德维大学（Shri Mata Vaishno Devi University）收集的《膝关节X线骨质疏松数据库》（Knee X-ray Osteoporosis Database）数据集，该数据集亦发布于Mendeley Data平台（链接：https://data.mendeley.com/datasets/fxjm8fb6mw/2）。该数据库包含350张膝关节X线影像，同样划分为正常、骨质疏松、骨量减少三类，但其类别分布存在不平衡问题：正常影像仅36张、骨量减少影像154张、骨质疏松影像49张。
由于两个数据集均存在类别不平衡问题，研究团队采用了多种数据增强技术以优化类别分布并提升数据多样性，具体方法包括随机水平翻转、随机垂直翻转以及不同角度的随机图像旋转。数据增强的目的是生成各分类下均衡的影像样本，同时提升深度学习模型的鲁棒性。经预处理与数据增强后，所有类别的样本数量均实现均衡。
最终合并后的数据集按照70%、20%、10%的比例划分为训练集、验证集与测试集，所制备的数据集将用于基于深度学习技术的膝关节放射影像骨质疏松症分类任务。