An assessment of the usefulness of image pre-processing for the classification of first trimester fetal heart ultrasound using convolutional neural networks

Abstract

Convolutional neural networks have obtained note-worthy results in various fields of medicine, giving an objective, virtual second opinion to the physicians concerning the presence of disease indicators. Usually, it is believed that the performance of the model arises when it can concentrate on the exact region of interest of the examination within the entire larger scan available. A prior processing of the medical images is thus usually helpful towards a better prediction by eliminating background information that is not of interest to the classification process. But what if this information is actually helpful when the problem needs the reference to an extra Other class that represents the samples containing nothing important from the detection point of view? Or what if the processing itself leads to artefacts misleading the model? In this respect, the paper investigates the effect of an image processing step to precede the deep learning model. The medical problem treated regards the identification of the four principal views in first trimester fetal echocardiography scans towards an early indication of possible congenital heart defects.