Özet:
The main objective of this study is to provide automatic recognition of IUGR in the early stages of pregnancy by using noninvasive method. The difficulty faced in interpretation of IUGR in the early stages forced researchers to study about automatic detection of growth restriction. We aim to make fast and effective classification of ultrasound readings that are collected from emergency deliveries. Using intelligent data analysis techniques, computer programs could easily interpret maternal, placental and fetal measurements, predict presence or absence of growth restriction and provide real-time analysis and diagnosis. In this study, several machine learning techniques have been applied to IUGR dataset for classification using PI (Pulsality Index), RI (Resistancy Index) of UA (Umblical Artery), MCA (Middle Cerebral Artery) and DV (Ductus Venosus), and AFI (Amniotic Fluid Index) measurements. These measurements are taken from ultrasound readings from the mothers at emergency room. After data acquisition and scaling processes of the data, we applied 13 different classification algorithms. These 13 classifiers that have been used in this study can be divided into three groups. First group consists of single classifiers such as Support Vector Machines, k-Nearest Neighbors and Logistic Regression. In the second group, we tried to reject low confident test instances to achieve higher classification accuracy with higher confidence. Third group uses hybrid classifiers in order to benefit from several classifiers. Among these groups, performance of second group outperformed the third and lowest performance obtained from the first group. Within second group, SVM classification with rejection of low confident test samples results are shown to outperform competing classification results.
