CA-SegNet: A channel-attention encoder–decoder network for histopathological image segmentation.

Histopathological image segmentation based on encoder–decoder architectures has emerged as a pivotal research area in medical image analysis. However, due to the irrelevant information within multi-channel representations from the encoder, the coarse reuse of shallow features in skip connections may burden the learning and even adversely affect the decoder. While various variants have been developed to cope with this issue, the performance remains unsatisfactory. In this work, we propose a novel encoder–decoder architecture named CA-SegNet to address the above issue more effectively and achieve advanced histopathological image segmentation. Our novelty is twofold: firstly, a bottleneck-structured decoder is developed to improve the integration of multi-channel feature representations, and secondly, a sequence of channel-attention feature fusion modules (CAFFMs) are developed to adaptively guide the reuse of fine-grained shallow features in skip connections while learning the channel-wise dependencies. Experimental results on different publicly available histopathological image datasets demonstrate that our CA-SegNet outperforms existing state-of-the-art methods on both large and small-scale datasets. • A novel deep learning-based CA-SegNet model for histopathological image segmentation. • A channel-attention feature fusion module to greatly improve shallow feature reuse. • A bottleneck-structured decoder for better feature integration. • Outstanding segmentation performance on both large and small-scale datasets. [ABSTRACT FROM AUTHOR]