Dilated Multi-Scale-Link Net with Merged Self Attention based Fetal Head Segmentation Using 2D Ultra Sound Image

Abstract

Automatic segmentation of fetal head from ultrasound images and head circumference (HC) biometric measurement remain difficultdue to the inherent characteristics of fetal ultrasound images at various stages of pregnancy. A quick and automated method of extracting quantitative measurements is desired to increase the efficiency of clinical work flows and support decision making. In the paper, we propose a new deep learning method for automatic fetal ultrasound image segmentation called Dilated Multi-Scale-LinkNet with Merged Self Attention(DMSL_MSA) based Fetal Head Segmentation.We evaluate the proposed model in the context ofsemantic segmentation of fetal head from 2D Ultrasound fetal image and compared the model with other state-of-the-art segmentation networks. The HC18 Grand Challenge dataset is used for model training and evaluation.It contains 2D ultrasound images at different pregnancy stages. The results of the experiments performed on ultrasound images of women in the various stages of pregnancy are presented. The proposed Dilated Multi-Scale-LinkNet with Merged Self Attention method was evaluated on the testing set of HC18 in terms of four performance indices: Dice similarity coefficient(DSC), Hausdorff distance (HD), HC difference (DF), and HC absolute difference (ADF).The results reveal that we achieved 96.37% Dice score, 1.35 mm ADF, 0.65 DF and 1.54 HD when segmenting the fetal skull.By incorporating the Merged Self Attention mechanism, the proposed method yielded better segmentation performance than the conventional Link-Net based methods.This demonstratesthe efficiency of our approach to generate precise and reliable automatic segmentations of 2D Ultrasound fetal images.