MSN-TCSeg: A Transcranial Sonography Dataset for Midbrain and Substantia Nigra Hyperechogenicity Segmentation

Obtain declaration Access to this dataset is available upon request. Researchers interested in obtaining the dataset should submit an access request to the designated contact person, including a brief description of the intended research purpose. Upon receiving the request, the contact person will consult with the relevant clinicians and project leaders. Access to the dataset will be granted after approval by the responsible clinicians and project leaders. For detailed information about the dataset, please refer to the associated publication: https://doi.org/10.1016/j.cmpb.2026.109605 Dataset Description This repository provides a curated transcranial sonography (TCS) segmentation dataset focusing on two anatomically and clinically relevant targets for Parkinson’s disease research: the midbrain and substantia nigra hyperechogenicity (SN+). The dataset is intended to support research on automated segmentation, benchmark evaluation, and methodological development in ultrasound-based neuroimaging. The dataset consists of two subsets: 1. Midbrain segmentation subset: 700 TCS images with corresponding midbrain segmentation masks. 2. SN+ segmentation subset: 370 TCS images with corresponding substantia nigra hyperechogenicity segmentation masks.In addition, probabilistic soft labels are provided for the SN+ segmentation subset as a supplementary reference. All images were acquired at the standard midbrain plane through the temporal acoustic window and were fully anonymized prior to release. Only images in which the midbrain contour was clearly identifiable were included, ensuring minimum annotatability and anatomical interpretability. Segmentation labels were generated following a rigorous double-blind annotation protocol by two experienced neurosonologists. Inter-observer agreement was quantitatively assessed on a per-case basis. Based on the observed agreement characteristics, target-specific gold standard construction strategies were adopted. For the midbrain, consensus labels were generated using automated label fusion followed by rapid quality review. For SN+, a stratified workflow was applied, combining automated fusion, non-blind re-evaluation, and expert arbitration for cases with high annotation uncertainty. All images and masks are provided in PNG format and organized to facilitate direct use in machine learning workflows, including cross-validation experiments. Usage Notes 1. This dataset is intended solely for non-commercial scientific research and educational purposes. 2. The data are not intended for clinical diagnosis or decision-making. 3. Users should cite the associated publication when using this dataset in academic work. 4. Redistribution of the dataset or any derivative versions should comply with the specified license. Ethics Statement The dataset was collected retrospectively under approval of the Ethics Committee of Guangzhou First People’s Hospital. All data were fully anonymized prior to analysis and public release. No personally identifiable information is included.