基于卷积神经网络的压力性损伤检测与识别方法研究及准确性分析OA

Objective:To construct an automatic identification model of stress injury staging based on convolutional neural network(CNN)to assist clinical nurses to improve the accuracy of staging judgment.Methods:The wound images of 354 patients with pressure injury from July 2022 to July 2023 in the First Affiliated Hospital of Guangxi Medical University were collected and marked by two senior wound therapists with reference to NPUAP staging standard.They were divided into training set(n=248)and test set(n=106)by stratified random sampling.The YOLOx-L target detection network is used for training and verification,and the robustness of the optimization model is enhanced by data.The evaluation inde-xes include classification accuracy,95%confidence interval(CI)and confusion matrix analysis(statistical Methods:Clop-per-Pearson CI,exact binomial test).Results:The overall recognition accuracy of the model was 58.59%(95%CI:48.60%～68.03%,P=0.041),which was significantly better than the random level(50%).The analysis of stage spe-cificity showed that YOLOx-L model had the best identification efficiency for stage 3 pressure injury(70.00%,95%CI:53.89%～82.84%,P=0.002),but limited identification for stage 1 and deep tissue injury(41.67%and 25.00%).The confusion matrix reveals the core error mode:the two-way misjudgment rate of stage 1 and stage 2 is 40%,and the missed diagnosis rate of deep tissue injury is 30%.Conclusion:YOLOx-L model has the ability to identify stage 3 pres-sure injury clinically,but it still needs to be optimized for early injury and atypical deep tissue injury.CNN technology can help improve the efficiency of staging judgment of nurses and provide technical support for early intervention of stress inju-ry.