Precise and reliable cell segmentation algorithms can be crucial for gaining insights into disease mechanisms, in diagnostic applications and for disease monitoring and thus therapeutic interventions. Deep learning (DL) has revolutionized medical image segmentation, offering accuracy that matches or surpasses human experts. However, DL’s effectiveness is constrained by its need for large volumes of expert-labelled data, incurring both time and expense costs for data preparation. Traditional cell segmentation techniques, while less dependent on expert data, typically are not very effective in the blurry border, or complex scenarios involving noise and diverse backgrounds. In order to address these challenges, we, firstly, propose a novel self-adaptive framework that combines the strengths of traditional segmentation and deep neural networks. Traditional segmentation models provide initial segmentation results to guide deep neural networks in learning segmentation features. This approach reduces reliance on extensive expert annotations, as experts only need to select a few segmentation results instead of manually segmenting cells. Additionally, it improves robustness in the face of image complexities. Secondly, we propose an Attention Gate U-Net (AGU-Net) model to enhance cell border delineation. Our AGU-Net effectively handles background complexity, cell heterogeneity, and contrast variations, significantly advancing the accuracy of segmentation in complex morphological scenarious which are common in histopathological images. The proposed framework was tested on in-house as well as public datasets. Our framework achieves 82% accuracy in cell detection, 97% IoU in cell segmentation for the in-house dataset, and 84% Dice in cell segmentation for the public dataset. Therefore, this framework provides a possible application for medical imaging segmentation.