Study Of Automatic Method For Placental Maturity Evaluation

The accuracy of placental maturity staging is important to the clinical diagnosis of small gestational age,stillbirth and fetal death.Fetal viabilities,various gestational ages,and complicated imaging process have made placental maturity evaluation a tedious and time-consuming task.Despite numerous developments of tools and techniques to access placental maturity,automatic placental maturity remains challenging.To address this issue,we proposed to automatically grade placental maturity by obtaining gray scale features from B-mode ultrasound and vascular blood flow information from color Doppler energy(CDE)images based on a discriminative learning architecture.Then,we applied an improved descriptor using a coarse-to-fine scale representation with scale-invariant feature transforming and gray feature for visual features extraction after key-point detection with difference of Gassian.Also,feature encoding with fisher vector was used to improve the performance.Finally,we used SVM to do classification,and got 92.7%of accuracy.Experimental results of different key point detectors,feature descriptors,encoding methods demonstrated that our method achieves promising performance in placental maturity evaluation and would be beneficial in the clinical application.The development of deep learning provided more opportunities for better accuracy.Due to limited feature descriptors cannot illustrate images completely,we proposed to apply the popular convolutional neural network technique to our task.Convolutional layers were used to extract features,and pooling layers were used to compute speed up.The end to end method can reduce artificial interference on feature selection for better performance.Data obtained earlier was used to training network,the later was used to test the pre-trained network.We did some experiments based on AlexNet,VGG-F,VGG-S,VGG-M,VGG-VD-16 and VGG-VD-19.The excellent results provide more opportunities for more appropriate network.In the traditional machine learning method,we proposed to use feature fusion and multi-layer Fisher encoding to improve the accuracy.Later,we tried to apply popular deep learning technique to placental maturity evaluation.This paper may provide other possibility for the clinical application.