DualSRA-Net: A Dual Squeezed Residual Network with Attentive Multi-scale Featuring for Different Medical Images Segmentation Tasks

Automatic image segmentation is crucial for computer-assisted diagnosis. The U-Net-based models have attained the highest cutting-edge performance in the field of medical image segmentation (MIS). However, most existing U-Net-based approaches have the issue of coarse segmentation of lesions. For different MIS tasks, we propose a highly efficient dual encoder-decoder network (DualSRA-Net). The proposed DualSRA-Net exploits the importance of residual, attentive, and squeezing mechanisms in two serial and different encoder-decoder subnetworks. In order to achieve high accuracy, these subnetworks utilize atrous spatial pyramid pooling (ASPP), squeezed residual blocks in different versions, squeeze-and-excitation (S&E), stem block, and attention blocks. ASPP extracts more context data while preserving the feature map's resolution. Hence, the model can capture coarse-finetuning features effectively at various semantic scales which enhances the quality of feature learning. The residual blocks enable the network to be substantially deeper, while S&E units increase the network sensitivity to important features while suppressing unnecessary ones. Moreover, the attention technique is introduced to add more focus on essential features, while stem block is used to widen the receptive field and maintain image information without significantly increasing computational cost. In order to evaluate the proposed DualSRA-Net, eight MIS datasets are employed including CVC-ClinicDB, Kvasir-SEG, ETIS-LaribPolypDB, CVC-ColonDB, and Kvasir-Instrument from endoscopic images, data science bowl (DSB 2018) challenge as a nucleus segmentation from microscopic images, and international skin imaging collaboration (ISIC) 2017, 2018 as skin lesions segmentation from dermoscopic images. Comprehensive experiments are performed, and the results are compared to other various methods. According to visual and computitative comparisons, the proposed network outperformed different techniques in terms of DSC, accuracy, recall, IoU, and precision.