A hybrid artificial intelligence (AI) model trained to analyze knee radiographs achieved 86% accuracy and 95% specificity for detecting osteoporosis, outperforming several existing deep-learning systems in a head-to-head comparison.
For the study, published by PLOS One, Korean researchers evaluated a new model, BONE-Net, designed to identify osteoporosis from standard knee X-rays. The model combines three components: a convolutional neural network (DenseNet169) that identifies local bone features; a transformer-based network (Vision Transformer) that captures broader structural patterns; and a custom “attention” module designed to highlight areas of clinical relevance, such as cortical thinning and trabecular disruption. These components feed into a final classifier that determines whether the radiograph is osteoporotic or normal.
The researchers used a publicly available dataset of 372 knee radiographs to train and test the system. Of these, 186 were osteoporotic and 186 were normal. They cropped the images to focus on relevant bone regions and resized them for analysis. Of the total images, 300 were used for model development, and 72 previously unseen images were reserved for final testing. To improve reliability, the training images were augmented through rotation, zooming, and cropping.
On the independent test set of 72 images, BONE-Net achieved an accuracy of 86.1%, a specificity of 94.7%, a sensitivity of 82.9%, precision of precision: 92.9%, a false-positive rate: 5.3%, and an area under the curve (AUC) of 0.92. Meanwhile, the confusion matrix showed 36 true positives, 26 true negatives, 2 false positives, and 8 false negatives.
In a separate part of the study, the researchers compared BONE-Net with other deep-learning models retrained on the same dataset, including VGG19, ResNet50, InceptionV3, EfficientNetB0, ConvNeXt-V2, DenseNet169, RDNet, and MetaFormer. BONE-Net demonstrated the highest overall accuracy and the lowest false-positive rate.
When compared with KONet, an ensemble model for knee osteoporosis detection published in 2024, BONE-Net showed better accuracy (86.1% vs 80.3%, respectively), specificity (94.7% vs 75.6%), precision (92.9% vs 71.1%), and a lower false-positive rate: (5.3% vs 24.4%).
“The enhanced performance of BONE-Net suggests significant potential for accurate osteoporosis detection, crucial for timely intervention to prevent fractures and improve patient outcomes,” the researchers wrote. “The model’s high accuracy, sensitivity, and specificity ensures reliable identification of osteoporosis, making it a valuable tool for clinical applications. This could assist healthcare professionals in making more informed decisions, ultimately reducing the incidence and impact of osteoporotic fractures.”
The investigators acknowledged several limitations of the analysis, including the relatively small dataset size, the fact that it focused on the knee only, and that it did not incorporate clinical variables such as age, sex, or bone mineral density.
They recommended that future work expand to other anatomical sites affected by osteoporosis, such as the hip and spine. “Additionally, incorporating multi-modal data such as clinical history, patient demographics, and BMD measurements into the model could provide a more comprehensive assessment of a patient’s osteoporosis risk,” they wrote.
The work was supported by the National Research Foundation of Korea grant funded by the Korea government. The researchers reported having no relevant conflicts.