Biomedical Engineering International Conference · Oct 2026
Performance Comparison Between Multi-Class and Sequential Binary Deep Learning Frameworks for Osteoporosis and Osteopenia X-Ray
By Pathomnop P, Waraporn P, Thadson K et al.
Osteoporosis and osteopenia are progressive bone disorders characterized by reduced bone density and structural deterioration, increasing the risk of fractures and highlighting the importance of early diagnosis. Existing deep learning approaches for knee osteoporosis screening often evaluate only a limited set of convolutional neural network (CNN) architectures, offering limited insight into how network design affects classification performance. Therefore, this study systematically compares multiple CNN architectures using both multi-class and sequential binary classification frameworks for automated knee X-ray classification of normal, osteopenia, and osteoporosis cases. The dataset consisted of balanced knee X-ray images from three classes and was randomly split into 80% for training and 20% for validation. Image preprocessing included resizing all images to 224 × 224 pixels, applying random horizontal flipping and random rotation to the training images, and normalizing the input data using ImageNet statistics. Eleven CNN architectures, including GoogLeNet, ResNet18, ResNet50, ResNet101, EfficientNet-B0, DarkNet19, DarkNet53, ShuffleNet, DenseNet201, VGG16, and MobileNetV2, were initialized with ImageNet pre-trained weights and evaluated using two classification frameworks: sequential binary classification (normal-versus-abnormal followed by osteopenia-versus-osteoporosis) and direct multi-class classification. Among the evaluated architectures, ShuffleNet achieved the highest performance for the normal-versus-abnormal classification task, with 94.76% accuracy, 95.04% precision, 94.73% recall, and a 94.74% F1-score. For the osteopenia-versus-osteoporosis classification task, DarkNet19 performed best, with 87.76% accuracy, 88.44% precision, 88.22% recall, and an F1-score of 87.76%. In the direct multi-class classification framework, VGG16 achieved the highest performance, with 79.53% accuracy, 80.17% precision, 79.62% recall, and a 79.59% F1-score. Overall, the sequential binary classification framework outperformed the direct multi-class classification approach, demonstrating the effectiveness of decomposing the three-class classification task into two binary classification stages.
doi.org/10.1109/bmeicon70815.2026.11773005Open accessBiomarkers