Post-scan ultrasound images are masked and labeled with anatomical features identified in the images. Masks corresponding to the features are transformed into the coordinate system of the pre-scan, raw ultrasound data frames corresponding to the images. The masks and their respective raw ultrasound data frames are used to train an artificial intelligence model. The model is then used to predict the presence of anatomical features of the same type in other pre-scan ultrasound data. Once predicted, the raw data is scan converted for display as segmented ultrasound images that identify the predicted anatomical features. Deployment of the model on pre-scan data avoids the need to account for different levels of processing that may be applied to post-scan images, while the labeling for training the model may be done on the post-scan images.