Detection and classification of fetal gender from ultrasound images based on deep learning techniques

All of the medical methods used in determining the sex of the fetus require expert intervention. In this study, a new method based on deep transfer learning algorithms is proposed by which fetal gender can be diagnosed independently from an expert. For the study, a data set consisting of 2-D, B-Mode obstetric ultrasonography images containing 4400 fetuses, 2200 female and 2200 male gender, was created together with the experts. In the first phase of the study, it is aimed to find the most compatible feature extractor network with the data set. By applying deep transfer learning techniques to VGG16, InceptionV3, ResNet152V2, DenseNet201 and Xception networks, new precision tuned models have been obtained. By classifying these models, the most successful classifier with an accuracy of 0.9627, ft-DenseNet201 network was chosen as the best feature extractor. In the second stage, it is aimed to find the best classifier. In the realization of this stage, feature extraction process has been performed with the convolution base of ft-DenseNet201 network and these features are classified with the algorithms of Logistic Regression (LR), Linear Support Vector Machine (LSVM), K-NearestNeighbor (KNN), Decision Tree (DT), Random Forest (RF) and AdaBoost (AB). Among the 11 different classifier algorithms used, the most successful model is ftDenseNet201 + LSVM with 0.9782 test accuracy. The proposed method can be integrated into household ultrasound devices so that parents can find out the sex of the fetus without a doctor's consultation. At the same time, the proposed system can be considered as a new diagnostic method, an alternative to the non-invasive obstetric ultrasonography method, which is the most commonly used method in gender diagnosis, and can provide opportunities that alleviate the workload of doctors.