Fig. 6: Visualization of ultraFedFM's pre-training, data augmentation, and segmentation performance.
From: From pretraining to privacy: federated ultrasound foundation model with self-supervised learning
a The reconstructed ultrasound images from the pre-trained model, where the masked regions are selected based on texture information. b To increase the richness and balance of features, images captured in linear-array mode and convex-array mode are transformed into each other. c Visualization of multi-class organ segmentation and the prediction of the angle of progression (AoP). d Visualization of binary lesion segmentation. Heatmaps highlight the attention areas of the features extracted from the pre-trained encoder. The closer the color is to red, the more the model pays attention to the area.
