TEM-NeutroStruct: A Robust Dataset of Neutrophil Ultrastructures for Deep Learning and Biomedical Applications

TEM-NeutroStruct is a rigorously annotated and validated dataset of neutrophil ultrastructures derived from transmission electron microscopy (TEM) images. Developed over 1 year and 9 months, this dataset provides high-resolution TEM images accompanied by high-quality annotations for object detection and segmentation tasks. It is the first publicly available dataset of its kind, designed to support research in automated TEM image analysis, hematological studies, and deep learning.

The dataset is available in two versions: one with four classes and another with seven classes. The four-class version reflects the original annotation structure, while the seven-class version was created to provide a more detailed classification of granule types, enabling more nuanced analysis and experimentation. Both versions are included to ensure transparency and flexibility, allowing researchers to choose the format that best suits their requirements or to develop alternative class division strategies.

The dataset includes:

83 High-Resolution training images and 10 test images.

21,143 annotations in the training set and 3,150 annotations in the test set.

Seven annotated classes: Non-specific Primary Granules, Non-specific Secondary Granules, Specific Granules, Non-specific Tertiary Granules, Empty Vesicles, Emptying Vesicles, and Vacuoles.

AnnotationsFormats Provided:

COCO JSON format: Compatible with object detection frameworks like YOLO, Faster R-CNN, and EfficientDet.

CVAT XML format: Retains the original elliptical shapes for segmentation and modification in annotation tools. These formats are easily convertible to other formats required for deep learning models.

Segmentation Masks: Masks are provided in PNG format, with each class represented by a unique color.

Dataset FeaturesImage Resolution:

87 images are 4096 x 4224 pixels.

6 images have a resolution of 2048 x 2115 pixels.

Classes and Annotations:

Seven biologically relevant classes are annotated to capture the structural diversity of neutrophils.

Each annotation has been rigorously validated for quality and accuracy.

Dataset StructureThe dataset is organized into two versions: one with the original four classes and another with extended seven-class annotations. Visit the README.txt file For more details on how the data is structured.  

Challenges and InsightsClass Imbalance: Non-specific Tertiary Granules dominate the dataset, while Vacuoles and Specific Granules are underrepresented.

Size Variability: Granule sizes vary significantly across images, necessitating adaptive classification techniques.

Intensity Variability: Classification of light vs. dense granules depends on nucleus intensity, which varies between images.

ApplicationsThe dataset is suitable for:

Benchmarking advanced computer vision algorithms for object detection and segmentation.

Research in hematology and clinical diagnostics.

Automated analysis of neutrophil ultrastructures to enhance understanding of neutrophil biology.

LicenseThis dataset is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0).  You are free to share, adapt, and use the dataset for any purpose, provided proper attribution is given.  For details, see the LICENSE.txt file.

CitationThis section will get updated once the data paper get published.