SegMed-Net: a domain-aware dual encoder-decoder network with dense feature fusion for enhanced medical image segmentation

Medical image segmentation is crucial in healthcare for identifying and delineating structures or abnormalities. This paper introduces SegMed-Net, a new architecture combining a dual encoder network with a dense aggregation decoder. The dual encoder integrates a pretrained model branch and a domain-specific feature branch, merging the outputs to address feature disparities. Reverse decoder blocks (RD-blocks) and dense aggregation decoders extract final segmentation feature maps. Evaluated on Kvasir-SEG, CVC ClinicDB, 2018 Data Science Bowl, and ISIC-2018 datasets using Jaccard Index (JC), Dice Coefficient (DSC), precision, and recall metrics, SegMed-Net outperforms state-of-the-art methods on three datasets. It achieves DSC scores of 91.4%, 90.38%, and 90.95% on CVC ClinicDB, Kvasir-SEG, and ISIC-2018, respectively, and JC scores of 86.4%, 83.7%, and 84.34% on the same datasets, demonstrating its superior segmentation capabilities.