ShuffleLeNet architecture for spinal cord injury classification and level detection.

PURPOSE: To develop an automated and accurate framework for classifying spinal cord injuries (SCI) and detecting injury levels from CT images, thereby enabling early diagnosis and timely clinical intervention while addressing challenges related to noise sensitivity, segmentation accuracy, feature extraction, and computational efficiency. METHODS: The proposed framework consists of multiple stages. Initially, Wiener filtering is employed for image preprocessing to remove noise and enhance CT image quality. A modified U-Net architecture is then utilized for accurate spinal cord segmentation. From the segmented images, multiple discriminative features are extracted, including Improved Local Gradient Increasing Pattern (ImLGIP), deep features from ResNet and VGG16, and PHOG features. Finally, a hybrid Improved ShuffleLeNet model, integrating the strengths of Improved ShuffleNet and LeNet, is used for SCI classification as well as injury level detection. RESULTS: The proposed Improved ShuffleLeNet model demonstrated superior performance compared with conventional methods. It achieved the highest classification accuracy of 0.962, a negative predictive value (NPV) of 0.972, and a precision of 0.935, indicating its effectiveness in accurately classifying spinal cord injuries and identifying injury levels. CONCLUSION: The proposed framework provides a robust and efficient solution for automated SCI diagnosis from CT images. By integrating effective preprocessing, precise segmentation, comprehensive feature extraction, and hybrid deep learning-based classification, the model enhances diagnostic accuracy while maintaining computational efficiency. The promising performance of the Improved ShuffleLeNet model highlights its potential for real-time clinical applications in spinal cord injury classification and injury level detection.