CAAGP: Rethinking channel attention with adaptive global pooling for liver tumor segmentation. (November 2021)

• Record Type: Journal Article
• Authors: Zhang, Chi; Lu, Jingben; Yang, Luxi; Li, Chunguo
• Is Part Of: Computers in biology and medicine. Volume 138(2021)
• Publication Date: 2021-11
• Subjects: Channel attention -- Self-attention -- Adaptive global pooling -- Liver tumor segmentation; Medicine -- Data processing -- Periodicals; Biology -- Data processing -- Periodicals; 610.285
• DOI: 10.1016/j.compbiomed.2021.104875
• Languages: English
• ISSNs: 0010-4825

Abstract

Abstract: Channel attention, a channel-wise method often used in computer vision tasks, including liver tumor segmentation tasks, is able to model the channel relationship to augment the representation ability of feature maps. Channel attention could adaptively generate channel-wise responses using global pooling, which aggregates spatial information roughly. Actually, global pooling may introduce the loss of fine information, which is vital for segmentation tasks. Hence, we rethink the problem and propose the channel attention with adaptive global pooling(short for CAAGP), which preserves spatial and fine-grained information for liver tumor segmentation tasks when channel attention is generated. The model consists of three main parts, including improved self-attention, adaptive global pooling and responses generation modules. Self-attention achieves excellent performance in the computing of the spatial attention, while introducing serious calculation and memory burdens. In order to remedy these burdens, we improve self-attention and consider aggregating spatial information from x and y directions respectively. Extensive experiments have been conducted to verify the effectiveness of our proposed method. Our CAAGP outperforms other attention mechanisms significantly in liver tumor segmentation, especially for tumors with small size. Graphical abstract: Image 1 Highlights: We propose an efficient and effective method named CAAGP, which could simultaneously aggregate the spatial information into channel attention. We improve the self-attention, which could effectively aggregate spatial information from x and y directions respectively with little calculation cost. Extensive experiments have been conducted to verify the effectiveness of our proposed method.