Large annotated datasets are essential to fully utilize deep learning methods in image processing. However, when it comes to medical imaging, such data is scarce due to privacy issues and the lack of experts for annotation. In conventional computer vision, generative models can synthesize images and automatically annotate them. Still, medical images pose unique challenges, where the complexity and variability in human tissue and disease structures make generalizing these methods difficult. We introduce a partition-based parallel latent diffusion model (LDM) to address this challenge. Medical images are divided into partitions based on grayscale similarity using image morphology operations. Each partition’s texture and shape are then used to fine-tune the LDM, creating an independent generator for each partition while preserving medical features. The synthetic partitions are subsequently combined and de-noised to produce a complete synthetic medical image with its corresponding segmentation mask. We applied this method to the automatic segmentation of glioma magnetic resonance imaging (MRI). We trained the segmentation models using synthetic MRIs generated from 2D slices of glioma MRIs, and the results show that U-Net trained on synthetic data generated by our proposed method achieves a relative DSC of 98.27% on the full dataset, and improves the DSC from a mean of 0.6324 to 0.7338 on datasets with only 16 MRIs, outperforming the state-of-the-art method trained on the original dataset.