Accurate detection of fresh and old vertebral compression fractures on CT images using ensemble YOLOR

Vertebral compression fractures resulting from osteoporosis contribute to pain and disability among older people, necessitating early detection and treatment. While MRI provides effective diagnosis, its higher cost poses a challenge. Conversely, CT imaging offers a more cost-effective alternative. However, CT-based detection of vertebral fractures may lack the accuracy of MRI. This paper presents a novel YOLO-based object detection method for localizing old and fresh fractures on spine CT images to expedite diagnosis and identify the optimal treatment window. We propose replacing the CSPDarknet53 backbone in the native YOLOR model with MobileViT and EfficientNet_NS, training three separate YOLOR models with different backbones. Subsequently, an ensemble approach is employed to leverage the enhanced feature extraction capabilities of the three models. Experimental results show individual model accuracies of 89%, 89.8%, and 89.2%, respectively. Furthermore, by replacing the convolution layer with the Involution layer and integrating the models using the ensemble method, we achieve a remarkable accuracy rate of 93.4%. The proposed Ensemble YOLOR model surpasses other state-of-the-art approaches, offering a fast and precise solution. This advancement provides valuable insights and guidance for physicians in diagnosing compression between old and new fractures. © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024.