U-Net Tuning Hyperparameter for Segmentation in Amniotic Fluid Ultrasonography Image

Ultrasound Amniotic Fluid (AF) images generally have image quality similar to other 2-D ultrasound images, which have noise, blurry edges, artifacts, and low contrast. Some of the confusing factors in pocket AF segmenting comprise (a) reverberation artifact, (b) AF mimicking region, (c) floating matters, and (d) incomplete or missing boundary. Obtaining the Region of Interest (ROI) area of amniotic fluid requires a segmentation method that can identify each object in more detail. Based on the problems in AF segmentation, the contribution of this research focuses on the development of segmentation methods in AF using the U-Net semantic segmentation model using the architecture of the Roonerberger. This paper analyzes several uses of hyperparameters to determine the performance of the U-NET model architecture, especially for segmenting AF. The hyperparameter tuning is in the optimizer, loss function, learning rate, and the number of epochs. The best performance of U-Net in segmenting amniotic fluid with a combination of RMSprop optimizer parameters, the Loss function is Binary cross entropy, learning rate value is 0.00001 with Epoch of 33 with DSC of 0.88 and IoU of 0.79, the accuracy of 0.87, precision of 0.93, recall of 0.88.