基于多模态深度学习的颈椎失稳症辅助诊断模型的初步构建与验证研究OA

Objective To preliminarily construct and validate a diagnostic model for Cervical Spine Instability(CSI)based on multimodal deep learning,and to develop an open-access,web-based prototype platform tailored for clinical scenarios.Methods Clinical data from 122 subjects were included,integrating three-position X-ray images and nine structured clinical variables.The model architecture featured an imaging branch built on ResNeXt50 and a clinical branch based on a Fully Con-nected Neural Network(FCNN),with cross-modal fusion performed at the feature level(ResNeXt50-FCNN).Modality ablation studies were conducted using two control groups:"Imaging-only"(ResNeXt50)and"Clinical-only"(FCNN).Evaluation met-rics included Area Under the Receiver Operating Characteristic Curve(AUC),Average Precision(AP),Accuracy,Precision,Recall,F1-score,and Specificity.Grad-CAM was employed for attention visualization of the imaging branch,while SHAP(SHapley Additive exPlanations)was used to interpretglobal and individual feature contributions of the clinical branch.Final-ly,the optimal model was deployed via the Streamlit framework to create an interactive,clinically usable platform.Results The multimodal model achieved superior comprehensive performance on the validation set compared to the control groups,yielding an AUC of 0.964,AP of 0.973,Accuracy of 0.919,Precision of 0.875,F1-score of 0.933,and Specificity of 0.813.Ablation studies demonstrated that the fusion strategy significantly compensated for the insufficient sensitivity of the"Imaging-only"model and the low specificity of the"Clinical-only"model.Grad-CAM visualizations revealed that the model consistently fo-cused on suspected abnormal structures across all three positions,effectively suppressing background interference.SHAP anal-ysis identified age,limitation of flexion-extension,tenderness,gender,and headache as the primary contributing variables.The developed AI platform(https://cervical-instability-classifier.streamlit.app/)provides real-time outputs of diagnostic probabili-ties and visualizations of feature contributions.Conclusion This study successfully constructed and preliminarily validated a CSI discrimination model based on multimodal deep learning,demonstrating promising diagnostic performance,interpretabili-ty,and potential for clinical application.However,given the limited sample size of this study,further external validation is re-quired to confirm the model's generalizability and ultimate clinical utility.