Automatic Measurement Of Fetal Head Circumference In Ultrasound Images Via Random Forests And Non-Iterative Fast Ellipse Fitting Method

Because of non-invasive nature, real-time acquisition, reduced cost, ultrasound imaging is the modality of choice in many clinical applications. Prenatal ultrasound examination can help reduce birth defect rates, which has great significance in reducing the economic and spiritual burden of the family and the society. Therefore, it is a hot research that how to get an efficient and accurate examination in clinic. In this examination, one of the most important tasks is to measure biometric parameters, including head circumference, which are the evidence to estimate fetal size, weight and to identify growth situation of fetuses.However, currently, measuring the biometric parameters manually is the dominant approach. The accuracy of this way heavily rely on the work experience of sonographers, besides, there are some difference between intra-observer and inter-observer. Thus, automatic measurement of biometric parameters has become a hot research. However, most of the automatic measurement methods of the fetal head circumference have some shortcomings, including running slowly,low accuracy, manually labeling the initial point in advance, which cannot be used in the actual clinical applications. Therefore, we propose an automatic, fast and accurate measurement method of the fetal head circumference. It is based on the Random Forests and the fast, non-iterative, geometric ellipse fitting method.To start with, the region of interest of head circumference was detected by the Random Forest algorithm which combined with the prior knowledge of medicine,so as to reduce the subsequent image region, and to eliminate the most interference of the noise and shadows. Next, we used a phase symmetry based method to detect the edge of head circumference, and this approach obtained the contour of the skull fast and effectively. Finally, we used a non-iterative based geometric ellipse fitting method to fit the ellipse of the skull.In this paper, our spot light is combining the Random Forest and prior knowledge to detect the ROI of the fetal head circumference. Compared with traditional Random Forest, it reduced the time cost of the detection, with an average about 3.7 times. In addition, we used a novel method, a non-iterative,geometric ellipse fitting method, to fit the ellipse of the fetal head circumference, and the advantage of this method is that the fitting part just cost 0.45 ms which is faster than the classical Hough Transform ellipse fitting method nearly1787 times. Last but not least, this fitting method could solve the problem of counter missing, which got a better result(96.66 ± 3.15% > 90.37 ± 9.03%) in Dice Similarity.We did some experiments on the 145 images to evaluate the proposed method.Experiment results showed that the Dice which compared with the sonographers.ground truth was 96.66 ± 3.15% and there was no significant difference(p value of t- test is 0.22), which justify our proposed method is effective.