In the image segmentation task, due to the complex distribution and variable shapes of brain tumors, conventional networks have poor accuracy in detecting small objects. Moreover, the computation cost of full convolutional network is high, and it is difficult to design a deep network as well. In this paper, we propose a new segmentation network called Attention-based Deep Residual U-Net (ADRU-Net). Res-Net50[1] replaces the fully convolutional network as the backbone feature extractor of U-Net[2], which can not only better learn deep features, but also avoid gradient disappearance. Both can greatly improve segmentation performance. At the same time, ADRU-Net introduces 6 Coord Attention modules[3] between the networks, which realizes the fine segmentation of edges without increasing the computational load too much. In practice, the introduction of Coordinate Attention not only optimizes the segmentation of tumor edges, but also makes the segmentation efficient especially in small objects. Experiments show that compared with other U-Net variant networks, ADRU-Net obviously improves the convergence speed, and can treat small objects well, which achieves a more effective segmentation with low cost. Finally, Our network proposed in this paper improves the mIoU by 5.13% and the mPA by 5.96% in the LGG Segmentation Dataset[4].
Attention-based Deep Residual U Neural Network on Brain Tumor Segmentation Algorithms
2022-10-12
1191722 byte
Conference paper
Electronic Resource
English
Computer generated hologram compression with attention-based deep convolutional neural network
British Library Conference Proceedings | 2021
|European Patent Office | 2023
|Efficient Lane Detection Technique Based on Lightweight Attention Deep Neural Network
DOAJ | 2022
|