Identifying pulmonary segments is essential for planning segmentectomy, a lung-conserving surgical procedure, but remains challenging since intersegmental boundaries are not directly visible in CT images and must be inferred from surrounding anatomical structures. To address this, we propose PSGM-TR, a transformer-based approach for pulmonary segment segmentation using Gaussian Mixture Models (GMMs). PSGM-TR regresses GMM parameters to construct GMM-parameterized spatial distributions, from which each location is assigned to the most probable segment. The GMM primitives that define each segment class are regressed from class-wise primitive queries, which are guided by features from 3D CT images and decoded through a GMM-parameterizing decoder. These queries implicitly learn semantically meaningful spatial roles and produce anatomically aligned primitives without relying on handcrafted anatomical priors. PSGM-TR achieves competitive performance and generates balanced, non-redundant, and anatomically coherent segmentations. Our analysis shows that PSGM-TR effectively captures anatomical structure through GMM-based spatial modeling, offering a clinically reliable and interpretable solution for segmentectomy planning and anatomy-aware image analysis.