MA-Net: A MLP-based Attentional Deep Network for Segmentation of Liver Tumor Ablation Region from 2D Ultrasound Image

Ultrasound image segmentation of ablation region of liver tumor has recently emerged as a significant tool for assessing tumor ablation surgery outcomes. However, existing segmentation methods are limited to the artifact of ultrasound images caused by hand-held transduces and speckle noise, leading to the background region likely being identified as the ablation region. Therefore, we introduce a MLP-based attentional network, MA-Net, and get accurate segmentation results. We present the hybrid attention cascading module to pay more attention to the ablation region to ensure accurate segmentation. In addition, we present an inverted residual multilayer perceptron module to avoid misrecognizing the ablation region as the background region. We evaluate our method on private and public dataset and achieve state-of-the-art performance.