S-ResNet-34: small sample-ResNet-34 for predicting cervical degeneration in x-ray image data.
Authors
Affiliations (6)
Affiliations (6)
- Department of Bone and Soft Tissue Oncology, Chongqing University Cancer Hospital, Chongqing, 400030, China.
- Chongqing Institute of Engineering, Chongqing Institute of Engineering, Chongqing, China.
- Department of Orthopedics, The Second Affiliated Hospital of Army Medical University, No.83 Xinqiao Main Street, Shapingba District, Chongqing, 400037, China. [email protected].
- Department of Bone and Soft Tissue Oncology, Chongqing University Cancer Hospital, Chongqing, 400030, China. [email protected].
- Department of Orthopedics, The Second Affiliated Hospital of Army Medical University, No.83 Xinqiao Main Street, Shapingba District, Chongqing, 400037, China. [email protected].
- Department of Orthopedics, The Second Affiliated Hospital of Army Medical University, No.83 Xinqiao Main Street, Shapingba District, Chongqing, 400037, China. [email protected].
Abstract
To construct an improved deep learning model that provides a more accurate and cost-effective solution for diagnosing abnormalities in the cervical physiological curvature. This study included 240 patients who were diagnosed with cervical spondyosis at our hospital from 2020-2024. Their X-ray images were collected and classified into three categories: normal cervical curvature, straightened cervical curvature, and reversed cervical curvature. The original image data were processed using the YOLO-V3 model for object detection, followed by data augmentation to generate an experimental dataset. On the basis of the ResNet-34 architecture, this study introduces a learnable weight matrix integrated with convolutional operations within the residual blocks to enhance the model's nonlinear representation capability, thereby leading to the construction of the proposed S-ResNet-34 model. Validation based on the test set indicated that the S-ResNet-34model achieved an accuracy (ACC) of 90.94%, an F1 score of 85.27%, and a recall rate of 85.36%. Compared with that of other models (CNN, SVM, RNN, DenseNet, and ResNet-152), the S-ResNet-34 model demonstrated superior performance in distinguishing cervical curvature abnormalities in small-sample X-ray image data. This study introduced a deep learned-based model, S-ResNet-34, which is an innovative approach to auxiliary diagnosis for small-sample X-ray image analysis. The findings offer a new choice for patients by simplifying the diagnostic process while maintaining diagnostic accuracy.