Deep Learning Based Detection and Classification of Amniotic Fluid Echogenicity Type for Enhanced Prenatal Diagnosis

Abstract

Meconium-stained and vernix caseosa are two significant components found in amniotic fluid, each serving distinct roles in fetal development. These components have potential to hinder fetal development when detected in excessive amounts. Due to the urgency of early detection, the development of a model for detecting and classifying echogenicity types is considered necessary. Detection model for echogenicity type based on digital images has not been developed previously. Therefore, In this study, we used a private dataset obtained from a gynecology clinic, consisting of 110 original images in JPG format. Our proposed model designed by applying a CNN based semantic segmentation approach to amniotic fluid images, with modifications made to the dense layer. Pretrained models incorporating various CNN architectures, are utilized to extract features from meconium-stain and vernix caseosa. Feature selection is carried out using three methods: Chi-Square, ANOVA, and Mutual Information. Xtreme Gradient Boosting algorithm for classification. Our proposed model achieve accuracy of 0.94 or 94% for classifying echogenicity type.