SECA-Net: Squeezed-and-excitated contextual attention network for medical image segmentation

•SECA-Net integrates SEMAC, SESCA, and MCFF into UNet, enhancing segmentation performance in medical imaging tasks.•SEMAC captures high-level and low-level features across multiple scales and receptive fields.•SESCA captures local and long-range dependencies by exploring the mutual dependence of spatial positions and channels.•MCFF block merges multi-scale features and reduces the semantic gap between the encoder and decoder.•SECA-Net achieves competitive segmentation performance across multiple public multi-modal datasets. Accurate medical image sgmentation is critical in the computer-aided diagnosis paradigm, serving as a crucial pathway for the prevention and treatment of various diseases. In this work, we present a novel network architecture SECA-Net for precise medical image segmentation across various modalities. SECA-Net integrates the traditional U-Net encoder-decoder structure with innovative modules to enhance contextual feature extraction and fusion. Specifically, we integrate a ResNet-based encoder into the backbone and design two key modules within the bottleneck: the squeezed-and-excitated multi-scale atrous convolution (SEMAC) module and the squeezed-and-excitated spatial-channel attention (SESCA) module. These two modules perceive deep features from encoder path and refine it to capture spatial-channel aware multi-scale features. Additionally, a multi-scale contextual feature fusion (MCFF) block is introduced to facilitate feature integration across encoder and decoder pathways. We assess the effectiveness of our proposed method across four public multi-modal datasets: BUSI, ISIC2016, Kvasir, and 2018DSB. The experimental results demonstrate the competitive segmentation performance of our proposed network compared to existing state-of-the-art models.