Synth-MS : A synthetic dataset for MS Lesion Segmentation

To address the challenge of limited annotated data and improve the accuracy of automated lesion segmentation in Multiple Sclerosis (MS), we generated a synthetic dataset comprising 80,000 brain MRI slices using our proposed Lesion-Guided Diffusion Network. This network is conditioned on lesion masks and lesion-specific features to produce realistic and anatomically consistent brain images with demyelinating lesions. Each synthetic image adheres to the applied conditioning, ensuring that the lesions exhibit plausible spatial distributions and morphological characteristics consistent with MS pathology.

The dataset also incorporates variability in lesion size, shape, and distribution through probabilistic lesion placement and morphological transformations applied during the Synthetic Mask Generation phase. This diversity enhances the dataset's utility for pre-training segmentation models, enabling them to generalize better to real-world data. The generated dataset represents a significant resource for advancing automated MS lesion segmentation and broader medical imaging applications, addressing the critical issue of data scarcity while maintaining clinical relevance.