Automatic detection and measurement of fetal femur length using a portable ultrasound device

摘要

A reliably estimated date of delivery (EDD) and gestational age (GA) are important to provide optimal care during pregnancy. Using ultrasound technology, the biparietal diameter (BPD) and femur length (FL) of a fetus are measured to calculate these parameters. Our research group has been developing a portable and user-friendly ultrasound scanner for the midwives inexperienced in the use of ultrasound in low- and middle-income countries (LMIC). The goal of this work was to develop an automatic method of detecting and measuring fetal femur length that can run on a tablet device to assist the health care worker during the scanning process. An ultrasound image containing a fetal femur was saved in the portable scanner. Colored UI elements were separated by selecting a threshold from HSV color space and the image was converted into grayscale. An adaptive binary threshold was selected by calculating the mean and standard deviation of the intensity of image pixels and a binary image was produced. Potential fetal femur contour candidates were computed from the edges between the black and white regions in the binary image. The fetal femur was detected by comparing its size, orientation, and position with other femur candidates. A Hough transform was applied in the location of the detected femur to find a straight line with the highest number of votes to measure the length of the fetal femur. All these steps were performed by using OpenCV image processing library. Fifty-eight different femur ultrasound images were acquired by ultrasound experienced (44 images) and inexperienced midwives (14 images). The gestational age range of the fetuses was between 18 - 32 weeks. The images had different quality, intensity, and zoom levels. The automatic method was able to detect femurs in 37 out of 44 images (84%) from experienced midwives, and 9 out of 14 images (64%) from inexperienced midwives. The automatic measurements were compared with the manual measurements from experienced midw- ves. The correlation plot and the error versus reference plot are presented. The correlation coefficient was R=0.95 and the mean error±1.96*STD was -2.34±6.52[mm]. Images containing curved femurs caused important FL underestimation; our method is currently being adapted to work on these cases. The average time of computation was 3.2 seconds in a Samsung P600 tablet device.