Identifying the nature of stomach diseases by ultrasonography based on genetic neural network

Clinical features and ultrasound signs of 76 subjects including 23 health subjects are collected. All samples are divided into the training set (38 samples) and the test set (38 samples) based on mean vector similarity. Each set contains 19 benign and 19 malignant. Multiple linear regression model (MLR-Model), back propagation neural network model (BPN-Model) and genetic algorithm-based back propagation neural network model (GABPN-Model) are established for distinguishing malignant from benign stomach diseases and are trained using the training set. Then three models are tested using the test set. The accuracy, the sensitivity and the specificity for the test set, GABPN-Model are 92.1%, 89.5% and 94.7%, BPN-Model are 89.5%, 89.5% and 89.5%, MLR-Model are 89.5%, 84.2% and 94.7%. Areas under curve of GABPN-Model, BPN-Model and MLR-Model are respectively 0.978, 0.945 and 0.958 in ROC analysis. These results confirm that Color Doppler ultrasound can be used as a tool for distinguishing benign from malignant stomach diseases (p < 0.01). Genetic algorithm-based neural network outperforms the multivariate linear regression and the back propagation neural network to establish a model identifying the nature of stomach diseases.