A Generalized Pooling for Brain Tumor Segmentation

Abstract

As a common malignant disease, brain tumor has high mortality. The automatic segmentation of brain tumor has significance for clinical diagnosis and surgery treatment. With the development of deep learning, CNN (Convolutional Neural Network) achieves remarkable performance in image processing and computer vision. Researchers have proposed a large number of CNN-based segmentation models such as FCN (Fully Convolutional Network) and Unet from the perspective of network architecture, loss function and attention mechanism. However, most of them are based on the traditional pooling operations such as average pooling and maximum pooling, which will lead to the loss of significant features or average features. Especially in brain tumor segmentation, tissues are usually quite small, so feature losing is more serious. More importantly, the fixed pooling patterns such as maximum pooling and average pooling, which cannot accommodate to varying data, may not be able to accurately express their features in down-sampling. In this study, we first unify maximum pooling and average pooling, and then propose a novel generalized pooling (GP) method with adaptive weights. This is the first work to improve models from the perspective of pooling operations for brain tumor segmentation. The experimental results show that our generalized pooling method is effective to segment brain tumors, outperforming the traditional pooling methods.

Keywords

• Average pooling
• brain tumor segmentation
• convolutional neural network
• generalized pooling
• maximum gpooling