IT-Block: Inverted Triangle Block embedded U-Net for Medical Image Segmentation

Xueyang Li, Yongfeng Huang, Cairong Yan, Lihao Liu

Abstract: Convolutional neural network (CNN) such as U-Net has demonstrated excellent performance for medical image segmentation. However, there are some limitations of its scalability. Specifically, the memory would increase significantly when embedding other functional modules into U-Net. Moreover, the kernel size used in U-Net is unitary, which makes it difficult to obtain the multi-level information and extract the target completely. In this paper, we only use 12 convolutional layers of U-Net as a backbone and design a novel architecture named Inverted Triangle (IT) Block embedded into it to address these problems. The IT-Block consists of Dense Connection, Residual Connection, and Inception, aiming to help the network obtain multi-level features and reuse them comprehensively. Furthermore, we optimize the dice loss to alleviate the butterfly effect, making the training process more stable during the backpropagation. The experimental results state that our framework is superior to U-Net in running time and accuracy.