Abstrak

Fetal biometric size such as abdominal circumference (AC) is used to predict fetal weight or gestational age in ultrasound images. The automatic biometric measurement can improve efficiency in the ultrasonography examination workflow. The unclear boundaries of the abdomen image and the speckle noise presence are the challenges for the automated AC measurement techniques. The main problem to improve the accuracy of the automatic AC segmentation is how to remove noise while retaining the boundary features of objects. In this paper, we proposed a hybrid ultrasound image denoising framework which was a combination of spatial-based filtering method and multiresolution based method.  In this technique, an ultrasound image was decomposed into subbands using wavelet transform. A thresholding technique and the anisotropic diffusion method were applied to the detail subbands, at the same time the bilateral filtering modified the approximation subband. The proposed denoising approach had the best performance in the edge preservation level and could improve the accuracy of the abdominal circumference segmentation.

This research aims to apply the localizing region-based active contour (LRAC) method to acquire the femur length in an ultrasound image automatically and to determine the effect of noise removal on the segmentation accuracy. The automatic femur length measurement system includes three main steps. The first step is the denoising process to reduce speckle noise in the ultrasound image. Afterwards, the LRAC method is applied to detect and segment a local region. The segmentation process with a certain number of iterations and a weight of the smoothing terms is started at the selected initial pixel. At the final step, the femur length is measured to estimate the gestational age. The experiment results show that the accuracy of the estimated gestational age increases significantly when the noise reduction technique is employed.

The research proposed a method that combined non-deep learning detector that called Aggregated Channel Features (ACF) detector and Convolutional Neural Network (CNN) that named Faster R-CNN detector to extract a cross-sectional area of the fetal limb in an ultrasound image. This combination is appropriate to solve the problem of object detection where the object has no clear characteristic, it has shape variation, blurred, and no clear boundaries, which is difficult to solve using the common thresholding or the edge detection method. This method also deals with the ultrasound image analysis which the training set is small. The pre-trained CNN can establish the classification model from the small annotated training data. ACF detector provides the region proposals of the non-cross-sectional area as an input of pre-trained CNN. The proposed method could improve the average precision of detection result when it was compared with Faster R-CNN and ACF detector alone. Also, the combination method could reduce the elapsed time of the Faster R-CNN training phase significantly.

No abstract