An ensemble deep learning framework for foetal plane identification

Antenatal (prenatal) care stipulates periodic monitoring of the foetus in alleviating risk factors and improving pregnancy outcomes. Foetal images generated from the ultrasound during prenatal care can be classified into different planes, such as the brain, abdomen, femur, etc., and can be further used for the abnormality detection, disease diagnosis, and foetal growth estimation of the foetus. Therefore, recognizing the foetal plane is one of the initial steps to automate antenatal care. This paper proposes an Ensemble Convolution Neural Network (ECNN) model that combines the base models to classify the foetal planes by using an open-access database containing 12,400 images with six foetal planes. Prior studies on this database involved state-of-the-art CNN methods, and Densenet-169 pre-trained model gave an accuracy of 93.6%. Three pre-trained CNN models (base learners) are trained using the transfer learning approach in the proposed method. The predicted features derived from the base learners are then used to train a Deep Neural Network (DNN) based meta-learner to achieve high classification rates. The stacked ensemble model resulted in an accuracy of \(96\%\) $96\%$ , which is better than the accuracy of the individual pre-trained models.