Globus Pallidus is an object in Magnetic Resonance Imaging (MRI) with severe intensity inhomogeneity and noises, such as small, low contrast, and weak boundary. Existing active contour methods fall into false boundaries while segmenting objects, which is smaller than the background like Globus Pallidus. This paper proposes a new local gaussian variational level set (NGVLS) for Globus Pallidus segmentation. We developed an energy term to create a smooth curve and prevent evolving curve into the false boundary. In the experiment, we compare NGVLS qualitatively and quantitively with existing methods such as Chan-Vese (CV), Region Scalable Fitting (RSF), Improved Region Scalable Fitting (Im-RSF), Local Pre-Fitting (LPF), and Local Gaussian Distribution Fitting (LGDF). We use Dice Similarity Coefficent (DSC) and Misclassification Error (ME) to measure the accuracy and error of the segmentation, respectively. The experiment is using 40 MRI datasets from Rumah Sakit National Hospital Surabaya, Indonesia. Qualitatively, the results show that NGVLS shows the best segmentation compared to the existing methods. Quantitatively, NGVLS achieves the highest average, minimum, and maximum value from DSC in 0.8291, 0.7119, and 0.9050, respectively. Also, NGVLS achieves the lowest average, minimum, and maximum value from ME in 0.0050, 0.0023, and 0.0155, respectively.