Application of Improved ResNet18 Based Neural Network for Non-invasive Blood Glucose Testing

Traditional invasive glucose testing carries a risk of wound infection and can cause significant discomfort to patients, while current optical methods are more convenient for practical use. Against this backdrop, this paper introduces a novel blood glucose concentration detection technique based on infrared thermal imaging. Firstly, a data augmentation method was proposed, and infrared thermal images were enhanced to address the issue of data imbalance. Then, the ResNet-18 convolutional neural network was combined with a support vector machine (SVM) and applied to the task of predicting blood glucose levels. Compared with the traditional ResNet-18 neural network and BP neural network, the algorithm proposed in this paper not only demonstrates greater stability but also superior performance, with a Root Mean Square Error and The absolute mean value of the error of 0.86 and 0.73, respectively. According to the Clarke Error Grid analysis, the results of our model solution fall within the acceptable range for clinical trials, indicating good potential for clinical application.