In the domain of neurological disorders such as Multiple Sclerosis (MS), MRI scans serve as pivotal tool for diagnosis and treatment evaluation. However, the availability of public datasets on Multiple Sclerosis is limited and insufficient. Although deep learning models have proven effective in aiding the diagnosis and analysis of disease progression, their accuracy is dependent on access to large and diverse datasets. To address this issue, we propose a novel approach by augmenting a UNet-based Denoising Diffusion Probabilistic Model (DDPM) with a spatial attention mechanism, benefiting from the interpretability of the generated attention maps to enhance the model’s convergence as well as the generated MRIs. Our modified architecture demonstrates superior performance, surpassing state-of-the-art GANs and diffusion models on existing public datasets as well as African Local dataset, with the lowest loss values 0.13 and 0.04 for L1 and L2 respectively, and a Fréchet inception distance FID of 0.20 between generated and ground truth scans. Showcasing its effectiveness in capturing the intricacies of MS lesions and the unique characteristics of MRI scans in the African context. Thereby broadening its applicability and relevance in diverse clinical settings. While the generated synthetic dataset offer a practical solution to the limited availability of real-world data and addresses the privacy concerns inherent in medical data sharing.