Optimizing TEM Image Segmentation: Advancements in DRU-Net Architecture with Dense Residual Connections and Attention Mechanisms

Abstract

This study introduces an innovative enhancement to the U-Net architecture, termed Modified DRU-Net, aiming to improve the segmentation of cell images in Transmission Electron Microscopy (TEM). Traditional U-Net models, while effective, often struggle to capture fine-grained details and preserve contextual information critical for accurate biomedical image segmentation. To overcome these challenges, Modified DRU-Net integrates dense residual connections and attention mechanisms into the U-Net framework. Dense connections enhance gradient flow and feature reuse, while residual connections mitigate the vanishing gradient problem, facilitating better model training. Attention blocks in the up-sampling path selectively focus on relevant features, boosting segmentation accuracy. Additionally, a combined loss function, merging focal loss and dice loss, addresses class imbalance and improves segmentation performance. Experimental results demonstrate that Modified DRU-Net significantly enhances performance metrics, underscoring its effectiveness in achieving detailed and accurate cell image segmentation in TEM images.