Image labeling and pretraining process
All images were labeled by a team of seven doctors with 3 to 23 years of
experience using LabelImg software (v. 2.0) following two steps. First,
five doctors with 3–8 years of experience identified lesions in the
images independently and labeled them with minimum bounding rectangles.
In addition, six normal structures were labeled if visible, including
cavity of septum pellucidum, thalamus, lateral ventricles, Sylvian
fissures, cerebellar and cisterna magna. Next, two senior independent
ultrasound specialists with over 20 years of experience verified the
labels for each image. After labeling, images from The First Affiliated
Hospital of Sun Yat-sen University and Dongguan M&C Health Hospital
were randomly assigned for training and evaluation with a ratio of 8:2.
The assignment was made on a case level rather than an image level,
ensuring that the testing dataset did not contain any images originating
from the training cases. Details are shown in Figure 1. To make the
algorithms robust, training datasets were augmented before training by
randomly rotating images from 0° to 60°, and flipping the images
horizontally and vertically to simulate various fetal positions.
Additionally, the images were zoomed up and down across the whole image
and were pseudo-color processed. After augmentation, all images were
resized to 1600 × 900.