Ultrasound liver tumor segmentation with nested U-Net and dynamic feature extraction

Abstract

Locating tumor from ultrasound images is of high importance in medical analysis and diagnosis. There are mainly two challenges for segmenting tumors from ultrasound images. The first challenge is the low resolution and high speckle noise of ultrasound images. The second challenge is the various shapes and sizes of tumors. We propose a multi-level feature extraction neural network to automatically segment the data. Our purposed model is trained and tested with liver tumor ultrasound images from an open source dataset. Specifically, after pre-processing the dataset with median filter and data augmentation, we employ a collaborative model that utilizes nested U-net, U-Net + + as a backbone. The model is integrated with dense short skip connections within sub-networks to further improve the gradient flow and feature preservation. In addition, we modify the original atrous spacial pyramid to an adaptive pooling for better compatibility with nested U-Net. Adaptive atrous spacial pyramid pooling is designed to extract features from different levels and cover the increasing range of feature extraction with regard to the depth of the nested network. Segmentation results showed that the proposed model outperformed multiple network structures, and achieved a 0.915 3 dice coefficient, 0.843 8 intersection-over-union and 0.0019 mean square error.