Attention based multi-scale nested network for biomedical image segmentation

Convolutional neural network-based methods have significantly enhanced the segmentation performance of biomedical images in recent years. Nevertheless, medical image segmentation presents a challenge marked by layout specificity, with limited variation between samples in medical datasets but significant variation within each individual sample. This aspect has been often overlooked by many models. Consequently, we propose a novel architecture called Attention based multi-scale nested network (AMNNet), specifically designed for efficient biomedical image segmentation. AMNNet comprises four components: early ReSidual U-CBAM (RSUC) modules and convolutional stages, a MLP stage in latent stage, and Convolutional Block Attention Modules (CBAM) integrated into the decoder stage. We introduce a lightweight CBAM to concentrate on regions proximate to the target and suppress extraneous features without substantial parameter increments. The RSUC module is proposed to combine receptive fields of different sizes, capturing comprehensive contextual information across various scales in medical samples. Extensive experiments conducted on the AMNNet reveal its outperformance compared to prevailing medical image segmentation methods across the ISIC2018, CVC-ClinicDB, CVC-ColonDB, BUSI, and GlaS datasets. Notably, AMNNet achieves Dice Similarity Coefficients (DSC) of 91.35%, 90.01%, 90.80%, 81.61%, and 94.31%, respectively.