Terabyte-scale supervised 3D training and benchmarking dataset of the mouse kidney

Area of Science:

• Biomedical Imaging
• Machine Learning
• Computational Biology

Background:

• Machine learning for 3D biomedical image segmentation underperforms compared to 2D image analysis.
• This performance gap is attributed to the scarcity of large, high-quality annotated training datasets.
• Acquiring such datasets requires advanced imaging, expert annotation, and substantial resources.

Purpose of the Study:

• To introduce the HR-Kidney dataset, a large-scale resource for 3D biomedical image analysis.
• To address the limitations of existing datasets in volume and quality for training machine learning models.
• To facilitate advancements in image processing, data augmentation, and machine learning techniques.

Main Methods:

• Acquisition of 1.7 TB of artefact-corrected synchrotron radiation-based X-ray phase-contrast microtomography images of whole mouse kidneys.
• Validated segmentation of 33,729 glomeruli.
• Inclusion of raw data, semi-automatic segmentations (renal vasculature, uriniferous tubules), and 3D manual annotations.

Main Results:

• Creation of the HR-Kidney dataset, a significant increase in data volume (1-2 orders of magnitude) over existing biomedical datasets.
• Provision of diverse data types including raw images, semi-automatic, and manual annotations.
• Establishment of a comprehensive resource for the scientific community.

Conclusions:

• The HR-Kidney dataset provides a foundational resource for advancing 3D biomedical image analysis.
• It enables further research in unsupervised, semi-supervised, transfer learning, and generative adversarial networks.
• This dataset is expected to spur innovation in machine learning applications within biomedical imaging.