A two-stage approach for the identification and classification of osteoporotic vertebral compression fractures

Osteoporotic vertebral compression fractures (OVCFs) are frequently underdiagnosed worldwide due to their subtle radiographic presentation and the inherently low contrast of x-ray images. This diagnostic challenge is further compounded by the limited availability of high-quality annotated datasets.

In this study, we proposed OVCFinder, a deep-learning-based two-stage cascade model designed for the accurate detection of OVCFs and the classification of old and new fractures in x-ray images.

The segmentation stage is built upon an enhanced DeepLabv3+ backbone, refined to improve local feature sensitivity by reducing the output stride, replacing atrous convolutions with standard ones to suppress noise, and incorporating multiscale convolutional layers to capture morphological variations. In the classification stage, a model library comprising nine diverse deep learning architectures supports both binary and three-class classification tasks, effectively distinguishing normal, new, and old fractures.

Experimental results demonstrate that the OVCFinder's performance exceeds that of single-stage detection frameworks. Furthermore, for binary classification, this cascade model obtained an average weighted accuracy of 62.72%, weighted precision of 72.90%, weighted recall of 62.72%, and weighted F1-score of 66.69%. The best individual model performances were achieved by AlexNet, with an accuracy of 70.49%, and VGG16, with an F1-score of 72.36%. These results notably exceeded the average performance of expert physicians (53.42% weighted accuracy, 53.51% weighted precision, 53.42% weighted recall, and 53.16% weighted F1 score). In the more challenging ternary classification task, this model achieved an average weighted accuracy of 37.05%, weighted precision of 50.16%, weighted recall of 22.91%, and weighted F1 score of 30.62%, with the upper-bound performance (ResNet152 reaching 48.09% accuracy and 39.27% F1 score) substantially surpassing that of human experts (39.74% weighted accuracy and 29.49% weighted F1 score).

This work represents the first application of a segmentation-classification cascade strategy for OVCF analysis, enabling end-to-end vertebral segmentation and fracture characterization directly on x-ray images. By decomposing the task, applying targeted optimizations, and adopting a modular design, OVCFinder achieves notable advantages in recognition accuracy, robustness, and interpretability. The OVCFinder's performance in both tasks demonstrates its clinical applicability and diagnostic superiority over conventional approaches, offering a low-cost, high-reliability auxiliary diagnostic tool for use in primary healthcare settings.