Deep learning-based automation for segmentation and biometric measurement of the gestational sac in ultrasound images.

INTRODUCTION

Monitoring the morphological features of the gestational sac (GS) and measuring the mean sac diameter (MSD) during early pregnancy are essential for predicting spontaneous miscarriage and estimating gestational age (GA). However, the manual process is labor-intensive and highly dependent on the sonographer's expertise. This study aims to develop an automated pipeline to assist sonographers in accurately segmenting the GS and estimating GA.

METHODS

A novel dataset of 500 ultrasound (US) scans, taken between 4 and 10 weeks of gestation, was prepared. Four widely used fully convolutional neural networks: UNet, UNet++, DeepLabV3, and ResUNet were modified by replacing their encoders with a pre-trained ResNet50. These models were trained and evaluated using 5-fold cross-validation to identify the optimal approach for GS segmentation. Subsequently, novel biometry was introduced to assess GA automatically, and the system's performance was compared with that of sonographers.

RESULTS

The ResUNet model demonstrated the best performance among the tested architectures, achieving mean Intersection over Union (IoU), Dice, Recall, and Precision values of 0.946, 0.978, 0.987, and 0.958, respectively. The discrepancy between the GA estimations provided by the sonographers and the biometry algorithm was measured at a Mean Absolute Error (MAE) of 0.07 weeks.

CONCLUSION

The proposed pipeline offers a precise and reliable alternative to conventional manual measurements for GS segmentation and GA estimation. Furthermore, its potential extends to segmenting and measuring other fetal components in future studies.