Automatic Detection of Standard Anatomical Planes of the Fetal Brain from 3D Ultrasound Volumes

Examination of the fetal brain during pregnancy is a crucial task to evaluate anatomical integrity and extract biometric data, which is essential for assessing fetal development. Traditional clinical practice for fetal brain analysis relies on the manual identification of standard anatomical planes from 2D ultrasound (US) imaging systems, which is a challenging task and highly dependent on the sonographer's expertise. Recent advances in 3D US technology have enabled the acquisition of volumetric data, facilitating the processing of 3D data to locate the standard anatomical planes within the volume. In this paper, a novel method to automatically extract standard anatomical planes of the fetal brain from 3D US volumes is proposed. The proposed method consists of a convolutional neural network (CNN) that estimates a probability map indicating plane location. Given that the standard plane presents specific structural features, particularly the presence of specific brain structures, a projection-based view containing image information is extracted from the probability map, allowing the network to learn the correct appearance of the view defined by the probability map. In the inference stage, a robust plane fitting is applied to the network's outputs to determine the final plane location. The proposed method was trained on 75 US volumes and evaluated on 40 US volumes, focusing on the detection of the transcerebellar plane. In terms of distance between centers and rotation angle between predicted and ground- truth planes, a detection error of 3.79 +/- 2.62mm/12.7 +/- 10.9 degrees was found. These results demonstrate that the proposed method can accurately detect standard anatomical planes, corroborating its potential to aid fetal brain assessment in clinical practice.