Smart Healthcare Approaches for Prenatal Anomaly Detection Using Machine Learning

Abstract

Prenatal surveillance of fetal health is important in low mortality rates by reducing both maternal and fetal mortality, especially in the underdeveloped areas where timely diagnosis is still unavailable. Cardiotocography and prenatal sonography are crucial in the context of determining the well-being of the fetus, but manual interpretation is very time-consuming and could be prone to inter-observer variation. This paper suggests the implementation of an automated framework of fetal anomaly detection based on deep learning and ensemble learning methods. A sample of 2,300 prenatal sonography images was obtained in one of the publicly available repositories on Kaggle and classified into three clinical relevant categories, which are healthy, benign finding, and malignant finding. Deep feature extraction was done using denseNet121 and stacking ensemble classifier based on the use of SVM, KNN, Random Forest and XGBoost as the base learners and a meta-learner to make the final prediction. Experimental tests done in Python and the Google colab platform with T4 GPU indicate an upward trend in performance with training epochs. The suggested model revealed a total accuracy of 92 percent, precision, recall and F1-score of 0.92, 0.92, as well as 0.91, respectively. The high recall demonstrates the good performance of the model in detecting real anomalies cases which makes the model a good decision support tool in early and reliable fetal anomaly screening.