Z-line is a junction between esophageal and gastric mucosa which is an important landmark in exploring esophageal diseases such as Gastroesophageal Reflux Diseases (GERD). This paper describes an effective interactive segmentation tool for Z-line annotation from Upper Gastrointestinal Endoscopy (UGIE) images. To this end, we propose a method containing of two main steps: firstly, a coarse scheme is designed to roughly segment boundary regions of Z-line. Thanks to recent advances of deep neural networks in biomedical imaging such as U-net segmentation, Z-line annotation is automatically achieved with acceptable results. However, the U-net’s segmentation is not accurate enough due to gastric mucosa complexity. We then propose a fine-tuning scheme, which aims to prune the U-net’s results. The proposed method is based on Binary Partition Tree (BPT) algorithms, which BPT is built-in into a Graphic User Interface. Objective of the proposed framework is to help endoscopy doctors achieve the best segmentation results with lowest efforts of interactions via the GUI. The experiment was setup to evaluate effectiveness of the proposed method by comparing performances of four different segmentation schemes. They are manual segmentation by hand, fully automation by U-net, the interactive segmentation via BPT only, and the proposed scheme (U-net+BPT). The results confirmed that the proposed method converged faster to ideal regions than the other three. It took the lowest time costs and users’ efforts but achieved the best accuracy. The proposed method also suggest a feasible solution for segmenting abnormal regions in UGIE images.