Improved CNN based on attention mechanism for sports dance fracture CT image recognition
Ganshu Li, BAT-OCHIR Magsar, D. Enkhtuya, Han Suo, Qian Wang, Panpan Deng · Journal of Radiation Research and Applied Sciences · 2025
In response to the low-recognition rate in computed tomography image recognition of sports dance fractures, an image recognition model based on attention mechanism and improved convolutional network is proposed. This model uses convolutional neural networks to automatically extract important features from computed tomography images, and achieves dynamic weighting and integration of features by introducing ResNeXt module, image space attention module, and feature space attention module structures. The improved model based on attention mechanism achieved an accuracy of 96 % on the testing set, which was about 5 % higher than that of the traditional convolutional neural network model. The loss rate decreased to 0.18, the mean square error value was 500, the structural similarity value was 0.82, and the brightness sequence error value was 289, all demonstrating its superior performance. In addition, the research method performed the best with a false detection rate of 5 %, which was lower than 5.0 %, 4.5 %, and 3.0 % compared with histogram equalization, demosaicing method based on deep learning, and global and local adaptive denoising network. The results show that the proposed method can significantly improve the recognition accuracy and reduce the false detection rate, which provides a new idea and method for medical image processing in sports. This study not only improves the accurate identification of fractures in athletes, providing important technical support for sports medicine, but also develops new perspectives for the application of medical imaging, highlighting the important connection between this particular field of sports dance and fracture medicine.