DenseRes-Net: An Architecture for Gland Segmentation

With the rise in popularity of deep learning techniques, the use of semantic segmentation algorithms have become increasingly attractive to. In this paper, we studied the U-Net architecture for its performance and quality of predicted masks on the GLAS dataset. The dataset contains images of haematoxylin and eosin (HE) stained slides and their respective masks. A new architecture called as DenseRes-Net has been proposed, which is an improvement over the existing U-Net. In the proposed architecture, we combine the strength of the DenseNet and the Residual Network to enhance the quality of feature representation. The proposed DenseRes-Net achieves a 0.89% higher F 1 score when compared with the U-Net. Additionally, we can observe a 2.69% increase in precision. These demonstrates the effectiveness of the proposed DenseRes-Net over the existing U-Net architecture.