Methods Of Automatic Measurement Of Fetal Femur In Ultrasound

In prenatal diagnosis,medical ultrasound imaging is the first choice because of its noninvasiveness,real time and low cost.The measurement of biometric parameters of fetal anatomy is an important part of prenatal diagnosis.At present,the biometric measurement is performed manually by clinicians,which may lead to a number of problems.First,the accuracy of the measurement is highly dependent on the experience of clinicians and has large variations.Second,accurate measurement is time-consuming.It may increase the time of diagnosis and make the efficiency of work flow low.Third,clinicians often suffered from repetitive strain injury.Therefore,we propose a method for automatic measurement of fetal ultrasonic biometric parameters,which aims to reduce the dependence on doctors’ experience and improve the diagnostic efficiency.First of all,related methods of biometric measurement of fetal ultrasound are surveyed and summarized.These methods are divided into image processing based methods,traditional machine learning methods and deep learning based methods.Because image processing based methods don’t need a lot images for training,and is easy to implement,we propose the automatic measurement method of femur length based on Frangi filter.However,the performance of image processing based methods is often influenced by the quality of ultrasound images,speckle noise and shadows.The incomplete boundary of fetal femur in ultrasound image often causes the relatively small measurement.Therefore,we propose a traditional machine learning method to detect the endpoints of fetal femur directly.,for the reason that it may avoid the error of unclear boundary.In traditional machine learning methods,we propose the automatic measurement method of fetal femur length based on random forest regression model.However,traditional machine learning method requires to manual design and select feature.,and these features need to be verified by several experiments.Therefore,we think deep learning methods which can ‘learn’ by itself.In deep learning methods,we propose the automatic measurement method of fetal femur length based on SegNet.,which not only can learn by itself but also has a better robust performance.Finally,we use 436 ultrasonic fetal femur images to evaluate the three methods above.The mean and standard deviation of the automatic measurement method of femur length based on Frangi filter is 3.33±7.83 mm,the automatic measurement method of femur length based on random forest regression model is 1.23±4.66 mm and the automatic measurement method of femur length based on SegNet is 0.46±2.82 mm.We also compared the results of the above three methods with doctor’s manual annotations,and it shows that the automatic measurement method of femur length based on SegNet has a best result of the three and robust.