An Advanced Cascaded UNet Architecture with Dual Attention for Bio-Medical Image Segmentation

This paper presents a novel approach to medical image segmentation by leveraging a double U-Net architecture. The first U -Net, with pre-trained V GGNet a sitsencoder and using ASPP and trainable decoder blocks, produces an initial segmentation. This initial segmentation enhances important features while suppressing irrelevant ones. The output is then fed into a MultiResUNet with dual attention mechanisms. These dual attention gates further refine the segmentation by s electively focusing on significant features a nd i gnoring t he lessrelevant o nes. The performance of the proposed model is evaluated using medical image datasets from various modalities, achieving satisfactory results. The proposed architecture significantly improves the precision in segmenting anatomical structures and pathological regions compared to popular and established U-Nets. These advancements in image segmentation technology hold promise for more reliable analyses in medical applications, potentially improving patient outcomes and aiding clinical decision-making.