Endometrial regions gives useful information to gynecologist for diagnosing or examining uterine, endometrium, and adnexal pathologies. To detect the endometrial lesions on TVUS, generally manual segmentation method is needed. It is labor-intensive and time-consuming. Moreover, delineating the border of endometrial region might be different from doctor to doctor due to the unclearness of boundary of TVUS image. Therefore, the automated and consistent guidance of finding endometrial region is beneficial. However, automated segmentation of endometrium is very challenging due to unclear boundary and heterogeneous texture. To tackle these issues, we propose a discriminator guided by endometrium key-point maps. The discriminator distinguishes a prediction map from a ground-truth segmentation map based on the key-point maps. The endometrium segmentation network strives to deceive the discriminator. In this adversarial way, the segmentation network can accurately find the boundary. Experimental results verify the performance of the proposed method on Samsung Medison dataset of the sagittal transvaginal ultrasound images.