Automatic Fetal Head Candidate Localization from 2D Ultrasound Images using Haar Cascade Classifier and Enhanced Localization Algorithm

Fetal head circumference (HC) is one of the fetal biometrics that is often used to determine gestational age and monitor the fetal growth in the womb. Nowadays, head circumference measurement from ultrasound images is performed manually by a doctor or sonographer by drawing a line or forming an ellipse to surround the fetal head. However, manual annotations are prone to human error and intra-observer as well as inter-observer variabilities. In this research, an automatic fetal head candidate localization was implemented using Haar Cascade Classifier (HCC) and further optimized by Enhanced Localization Algorithm (ELA). The combination of HCC and ELA was evaluated on 703 ultrasound images of the second trimester and 141 ultrasound images of the third trimester using the Jaccard Index (JI), Dice Similarity Coefficient (DSC), and Overlapped Area Ratio (OAR). The localization results showed that the HCC + ELA produced an average JI of 90.5%, DSC of 94.58%, OAR of 97.77% for the second trimester and an average JI of 88.17%, DSC of 93.33%, OAR of 96.97% for the third trimester. Based on the three evaluation parameters, we analyzed the factors affecting the accuracy of the localization algorithm and the correspondence of the localization results with the ellipse fitting outcome as the final process to determine the fetal head circumference.