A weighted bag of visual words model for predicting fetal growth restriction at an early stage

TL;DR

This study proposes a new model using placental ultrasound images to predict fetal growth restriction early, potentially helping doctors intervene sooner.

Contribution

The novel contribution is a weighted bag-of-visual-words model for early detection of fetal growth restriction using placental ultrasound images.

Findings

The proposed model achieved an accuracy of 70% and an F1 score of 0.7653 in predicting fetal growth restriction.

HOG feature extraction was found to be more suitable for placental ultrasound images compared to other methods.

The model's ROC curve had an AUC value of 0.80, indicating good predictive performance.

Abstract

Fetal growth restriction (FGR) is a significant concern for clinicians and pregnant women, as it is associated with increased fetal and neonatal mortality and morbidity. Although ultrasound has been the gold standard for many years to define FGR, it remains less than ideal for early detection of FGR. Placental dysfunction is a key factor in the development of FGR. The objective of this study is to achieve the early detection of FGR through the utilization of placental ultrasound images.

A retrospective analysis was conducted using 80 placental ultrasound images from 40 FGR fetuses and 40 normal fetuses matched for gestational age. Approximately 300 texture features were extracted from the placental images using key texture feature selection and histogram of oriented gradients (HOG) extraction methods. These features were then re-encoded using a bag-of-visual-words model with weight scaling, resulting in more effective features. The encoded image features were used to train a classifier, and ensemble prediction techniques were used to improve classification accuracy.

In this study, we applied the proposed method alongside several popular image classification methods for predicting FGR. The proposed method achieved the best experimental results, with an accuracy of 70% and an F1 score of 0.7653. We also compared different feature extraction methods separately, and the experimental results showed that HOG feature extraction is more suitable for feature extraction of ultrasound placental images. Finally, we plotted the receiver operating characteristic (ROC) curve with an area under the curve (AUC) value of 0.80.

To enable early prediction of FGR, we propose a visual bag-of-words model based on weight scaling for analyzing placental ultrasound images in the early stages—before significant fetal impairment occurs. The proposed model shows strong potential to assist doctors in making preliminary assessments, thereby facilitating earlier intervention. This can help reduce the risk of harm to both fetuses and pregnant women.

Figures