A channel attention segmentation model with cascaded multiscale convolutions

Abstract Deep learning-based methods have recently achieved high performance in medical image segmentation, and many variants of the U-Net have also been derived. Many variants have introduced attention mechanism methods in order to capture and utilize useful feature information more effectively. However, these attention mechanistic methods have weak non-local receptive field-enhanced connectivity for small objects, and key features of small objects may be overlooked, leading to a decline in segmentation performance. In addition, some variants add a simple cascade of multi-scale convolution modules, but only roughly pay attention to the feature maps of different scales, which are not sufficient for small object segmentation. So we propose a novel method with cascading multiscale convolution and channel attention mechanisms to meet the requirement for small object segmentation accuracy. Cascading multiscale convolutional modules in each hopping connection pathway are used in our approach to capturing multiscale features across different semantic depths. Furthermore, through the weight function, the channel attention module provides the importance weight for each channel for the small object feature. The experiments on the Prostate public dataset indicate that our method outperforms the previous models, reflecting the improved performance of small object segmentation.