Improving classifiaction performance of fetal umbilical cord using combination of SMOTE method and multiclassifier voting in imbalanced data and small dataset

The umbilical cord is one of the important organs on the growth and development of the fetus in the womb. Umbilical cords are associated with an adverse perinatal outcome such as intrauterine deaths, preterm delivery, repetitive intrapartum fetal heart deceleration, operative delivery for fetal distress, meconium staining, and chromosomal abnormalities. Initial screening stages of the fetal umbilical cord are carried out by analyzing the coiling pattern of two umbilical arteries. In this study, we propose a relevant feature extraction for classifying this organ based on texture and morphological approach. However, this study is facing an imbalanced class problem, which leads to the inability of the traditional classifier to predict data in the minority class. To deal with the emerging issues, this study proposed a model by optimizing data and algorithmic levels using a combination SMOTE method and Multiclassifier Voting. At the data level, the SMOTE method is used to generate new synthetic data and to balance the skewed data distributions directly. Subsequently, the classification uses a multiclassifier method that combines several traditional classifier methods in making final decisions based on voting schemes. The first experiment was conducted on imbalanced and small size data with a total of 62 umbilical cord images from 3 classes namely hypercoiling, hypocoiling, and normalcoiling. The results showed the multiclassifier voting method was able to achieve the best results with an accuracy of 73, average recall of 72, and ROC of 48 compared to other classifier methods such as SVM, Random Forest, KNN, Naive Bayes, and Decision Tree (C.45). However, all classifiers failed to predict the normocoiling class because of the limited amount of normocloing data in the trainning phase. Then the second experiment was carried out by adding synthetic data using the SMOTE method with the total data increasing to 111 images spread evenly in each class. The results show a combination of multiclassifier voting and SMOTE methods ultimately leading and produced higher performance than other classifiers, which yielded and accuracy of 81.4, average recall of 80, average precision of 81.5, and ROC of 89.1. © 2020, Intelligent Network and Systems Society.