Abstract

Brain magnetic resonance imaging (MRI) is widely used as a diagnostic and monitoring tool for most neurological conditions. Although clinical practice produces large volumes of data, they are typically underutilised for training machine learning models that could substantially improve the feasibility, accuracy, and implementation of diagnostic tools in clinical settings. Instead, such models are usually trained on small, highly curated datasets and evaluated under controlled conditions, which limits their generalisability to the variability of real-world clinical imaging. This limitation is also reflected in publicly available datasets. And, it is particularly evident in epilepsy-related abnormalities, for which such datasets are scarce.
The PMG-BrainDrugs dataset is the first open polymicrogyria (PMG) dataset including patients across the full age range. PMG is a malformation of cortical development characterised by abnormal cortical layering. The associated abnormalities are heterogeneous and can sometimes be difficult to detect. The dataset is comprised of 208 subjects, including 119 patients with radiologically confirmed PMG and 89 controls with negative findings. Each scanning session may include several anatomical imaging modalities (T1-, T2-, and T2*-weighted, or FLAIR sequences). For 27 subjects, 2 or 3 time points are available. The release of this dataset includes minimally processed data (with only defacing applied) as well as masks of PMG abnormalities. The dataset is available at DOI: