Structure-aware Loss Function for Ultrasound Image Segmentation

In deep learning, loss function plays a crucial role in training an effective neural network model. In the task of ultrasound image segmentation, the pixel-wise loss functions such as cross-entropy loss and dice loss are usually used to train a deep neural network model. These loss functions only count the distribution differences between the predicted results of the model and the ground-truth labels at pixel level, but do not pay attention to the consistence of the spatial relations between different tissues in the predicted results and the real images. In order to improve the semantic expression of the loss function in measuring the difference between the predicted results and the ground-truth labels, we use the concept of relative fuzzy connectedness to add the relative position relationship between tissues in ultrasound images to the loss function as a prior knowledge.