A method of pulmonary embolism segmentation from CTPA images based on U-net

The doctor's assessment of the degree for pulmonary embolism often needs to calculate the volume about it. The most important thing is to accurately segment pulmonary embolism in the Computed Tomography Pulmonary Angiography (CTPA) images. This paper proposes a method to segment the pulmonary embolism from CTPA images and the method is based on U-net which is an effective semantic segmentation network in deep learning. This work uses partial weights from the VGG16 pre-training model to initialize the parameters of the contracting path in the U-net. It can extremely reduce time of training and improve the generalization ability of the network. However, the class of the samples is unbalanced extremely. This paper defined a new loss function which is combined Focal loss and Dice loss to solve this problem. And the experiment shows that the proposed method can get effectively segmentation of pulmonary embolism in CTPA images.