Amniotic fluid segmentation based on pixel classification using local window information and distance angle pixel

The amniotic fluid surrounds and protects the fetus from colliding with one another during the uterus
development process. It also protects the umbilical cord from the uterine wall pressure, helps fetus
movement, and develops muscles and bones. Selection of the most profound areas of improper and withdrawal points are not straight caliper is very likely that affect the outcome screening. Furthermore, there are similarities in texture and gray level between objects, especially in the boundary area between amniotic fluid and other objects, such as the placenta and uterus, which causes the border
area to be less clear. Therefore, this research proposes a novel pixel classification model to separate
amniotic fluid from other objects with a limit on the specified window size to solve this issue.
In contrast to the most existing semantic segmentation methods or pixel-wise classification, we
use the sampling window technique to construct train sets of data to produce pixel-level information
more specifically in certain areas. Furthermore, each window extracts pixel information on gray level
features and local variance (GLLV), using a novel Distance Angle Pixel (DAP). To evaluate the proposed
model performance, we perform an extensive comparison with state-of-art methods by testing it on
amniotic fluid ultrasound images. The results showed that the proposed model with a 3 × 3 window
and random forest classifier could achieve the best value using an average Dice similarity coefficient
(DSC) of 0.876, Jaccard/IoU of 0.768, and Pixel accuracy of 85.7%. The proposed model has 0.324 DSC improved from U-Net, 0.046 from gray-level pixel classification, 0.092 from thresholding, 0,252 from active contour, and 0.19 from rectangle window sampling.