Cell nuclei segmentation is essential for microscopic image analysis. It facilitates the detailed micro-environmental insights for clinical studies. Automated nuclei segmentation can simplify the work of pathologists and address the variability and subjectivity among them, thereby improving diagnostic consistency and accuracy. Although deep learning (DL) techniques usually offer better performance than traditional methods for nuclei segmentation, they still struggle with challenges especially when the nuclei are clustered and overlapped with each other. To address these challenges, many effective deep learning techniques have been developed. Nonetheless, these approaches still exhibit limitations, such as the tendency to overlook certain nuclei. To mitigate this issue, we propose a two-stage network to boost the accuracy by incorporating an enhancement network (second stage) on top of the widely used encoder-decoder architectures (first stage). The enhancement network refines the results by utilizing the decoder’s output and original image. The primary goal of this proposed method is to segment those regions that might have been overlooked by the base model, by re-considering the original image within the network. The study employs popular models such as U-Net, Micro-Net, and U-Net++ as base models. The results illustrate improvement of \(0.8\pm 1.8 \%\) in Precision, \(1.5 \pm 3.0 \%\) in Recall, \(1.4 \pm 0.5 \%\) in Dice Score, and \(1.6 \pm 0.7 \%\) in IoU across different datasets through the proposed method.