Knowledge-based statistical data analysis for deep learning and voting classifiers merger

Abstract

Maternal and fetal morbidity and mortality can be prevented through careful monitoring during pregnancy. Using artificial intelligence doctors can establish a more accurate and a much faster diagnosis. In this paper, we propose a merger between different Deep Learning algorithms and three types of voting classifiers, unweighted and weighted hard voting, and soft voting for differentiating the abdomen view planes of a second trimester fetal morphology ultrasound. The deep learning algorithms that were applied on our dataset were ResNet50, InceptionV3, EfficientNetV2S, and MobileNet3Large, [19–22]. Our goal was to determine which voting classifier method better suits our data; hence we have made a quantitative comparison between unweighted and weighted hard voting applied on the committee of deep learning algorithms and soft voting applied on the same committee. We have also performed a thorough statistical benchmarking process to validate our findings. The obtained results showed that soft voting and hard voting with unequal weights outperform the vote of hard voting with equal weights classifier and the stand-alone algorithms.