Pulmonary embolism image segmentation based on an U-net method with CBAM attention mechanism

Abstract

Pulmonary embolism (PE) is a common lung disease that has attracted increasing interest in recent years because of its high morbidity and mortality. Doctors manually segment pulmonary embolism from CTPA images to evaluate the patient’s risk level. However, this task is time consuming and laborious, and misdiagnosis can easily occur by using traditional methods because of the small size of pulmonary embolism and the category imbalance between foreground and background in the image. In this paper, we use the classical medical image segmentation model U-net for our pulmonary embolism segmentation scenario, and make the following improvements based on it for the mentioned problems. (1) We use the CBAM attention mechanism to make the U-net focus on the key areas in the image. (2) We use a hybrid loss function that consists of Dice loss and Focal loss to improve the category imbalance problem of pulmonary embolism target being too small in CTPA images. The final experiment results showed that our method was able to achieve 0.87027 on the mean IOU and 0.93063 on the Dice Score, thus proving the effectiveness of our method in segmenting pulmonary embolism.