Dataset Card for SCC Real Brightfield Microscopy Cells (scc_real)

Dataset Summary

This dataset, named scc_real, consists of 5,000 real brightfield microscopy images of CHO cells (CHO-K1 and CHO DG44 variants). These images are patches of size 512x512 pixels extracted from larger images acquired using a high-throughput microscope. The dataset is intended to serve as a baseline for training and evaluating object detection models for single cell detection in brightfield microscopy. It contains manually verified bounding box annotations for single cells.

This dataset is used in the paper "Diffusion-Based Synthetic Brightfield Microscopy Images for Enhanced Single Cell Detection".

Supported Tasks

• Object Detection: The primary task is to detect single cells in brightfield microscopy images. The dataset provides bounding box annotations suitable for training object detection models to locate and classify single cells.

Languages

• Image Data: Primarily visual data representing brightfield microscopy images of cells.
• Documentation: English

Dataset Splits

• Train: 5,000 images
• Validation: 2,527 images (held-out real images, constant across experiments)
• Test: 16,758 images (held-out real images, constant across experiments)

The validation and test sets are held constant across all experiments described in the associated paper to ensure consistent and comparable evaluation of model performance.

Licensing Information

License: apach-2.0

This dataset is intended for research purposes. By using this dataset, you agree to cite the dataset and the associated paper in any publications.

Dataset Statistics

• Image Format: 8-bit grayscale PNG or JPG (specify the format you will upload)
• Image Size: 512x512 pixels
• Annotations: Bounding boxes in YOLO format (or COCO, specify the format you will use)
• Number of Classes: 1 (cell)
• Annotation Type: Manual and semi-automated. Real images were initially annotated using an algorithm based on fluorescence channel information and then manually verified and refined by expert human annotators.

Dataset Sources

• Microscopy Acquisition: CELLAVISTA 3.1 RS HE high-throughput microscope (Synentec GmbH), 10x objective lens.
• Cell Lines: Stably transfected CHO cells (CHO-K1 and CHO DG44 variants) expressing eGFP.
• Imaging Conditions: 96-well plates, three different cell densities (300, 150, and 1 cell per well).

The brightfield microscopy images were acquired as part of a research project focused on cell detection and analysis in high-throughput microscopy.

Data Fields

• image:
  • Type: Image (path to image file)
  • Description: Brightfield microscopy image patch of CHO cells. Grayscale 8-bit image data representing cellular structures and background.
• objects:
  • Type: Bounding Boxes
  • Description: Bounding box annotations for single cells in the image. Each annotation represents a single cell instance.
  • Format: class_id, bounding box coordinates and normalized bounding box coordinates.

Classes:

• cell: Single CHO cell

Considerations for Using the Data

Social Impact

This dataset is intended to promote research in automated microscopy image analysis, particularly for single cell detection. Accurate cell detection is crucial for biological research and drug discovery, enabling faster and more objective analysis of cellular phenotypes and behaviors. By providing a publicly available dataset, we aim to facilitate the development of robust and reliable algorithms, reducing the need for labor-intensive manual analysis in biological research.

Discussions of Bias

• Cell Type Specificity: The dataset is based exclusively on CHO cell lines (CHO-K1 and CHO DG44 variants). Models trained on this data might exhibit bias or reduced performance when applied to other cell types with different morphologies, sizes, or staining characteristics. Users should be aware of this cell-type specificity when applying models trained on this dataset to new biological contexts.
• Imaging Conditions: The data was acquired under specific brightfield microscopy settings and with a particular microscope model (CELLAVISTA 3.1 RS HE, 10x objective lens). Variations in microscope models, illumination settings, image acquisition parameters, and image processing pipelines in other labs might affect the generalizability of models trained on this dataset. Domain adaptation techniques may be necessary when applying these models to images from different sources.
• Annotation Bias: Although annotations were manually verified by experts with extensive experience in microscopy image analysis, there might still be inherent subjectivity and potential for human error in bounding box annotations. This is particularly true for densely packed cells or cells with weak contrast against the background. The quality of annotations is crucial for the performance of supervised learning models, and users should be aware of potential annotation noise.
• Well Edge Artifacts: While image patches containing visible well edges (dark arch artifacts) were manually excluded during dataset creation, subtle artifacts or non-uniform illumination related to well edges might still be present in some images, particularly near the image borders. These subtle artifacts could potentially introduce a subtle bias in model training, although efforts were made to minimize their presence.

Other Known Limitations

• Dataset Size: While 5,000 training images is a reasonable size for initial experimentation, larger datasets are generally beneficial for training deep learning models and could potentially lead to further improvements in model performance and generalization. Data augmentation techniques were used during model training in the associated paper to partially mitigate the limitations of dataset size.
• Image Patching: The use of image patches (512x512) extracted from larger raw images might result in the loss of some global contextual information that could be present in the original larger images. While patching was necessary to manage computational resources and focus on local cell features, future work could explore models that can process larger input images or incorporate global context.
• Limited Cell Density Variation: While images were acquired at three different cell densities (300, 150, and 1 cell per well), the range of cell densities might not fully cover all possible scenarios encountered in biological experiments. Datasets with a wider range of cell densities and cell arrangements could further enhance model robustness and applicability.

Additional Information

Dataset Creation Process

• Real Image Acquisition: Brightfield microscopy images of CHO cells (CHO-K1 and CHO DG44 variants expressing eGFP) were acquired using a CELLAVISTA 3.1 RS HE high-throughput microscope with a 10x objective lens, resulting in 8-bit grayscale images. Images were acquired from cells cultured in 96-well plates at three different seeding densities.
• Patchification: Raw subwell images (3056 x 3056 pixels) were divided into 512x512 pixel non-overlapping patches to increase the dataset size and focus the model on local cell features.
• Well Edge Filtering: Image patches containing visible well edges, identified as dark arch artifacts due to light refraction at the well boundaries, were manually excluded to ensure the training data primarily comprised relevant cellular structures.
• Dataset Sampling: A random subset of 10,000 patches was initially filtered and sampled for diffusion model training purposes (as described in the associated paper). From the filtered set of real patches, 5,000 patches were randomly selected to form the scc_real training dataset. Remaining real patches were allocated to validation and test sets.
• Annotation (Real Images): Real images in the training, validation, and test sets were labeled using a semi-automated approach. First, a classical image processing algorithm was implemented to automatically detect cells in the fluorescence channel (eGFP signal, where available). This was followed by manual verification and refinement of the automatically generated bounding boxes by a human expert using the Roboflow annotation tool to ensure high annotation quality.

Preprocessing

Images are provided as 8-bit grayscale PNG or JPG files.

Patches are non-overlapping 512x512 pixel crops from original microscopy images.

No further preprocessing steps were applied to the images in this dataset. Users may choose to apply additional preprocessing steps depending on their specific modeling requirements.

Maintenance

This dataset is a static release and will not be actively updated. For any questions, suggestions, or bug reports related to the dataset, please use the "Discussions" tab on the dataset page on the Hugging Face Hub.

Community Contributions For questions, suggestions, or bug reports related to the dataset, please use the "Discussions" tab on the dataset page on the Hugging Face Hub. We encourage community contributions to improve and expand this dataset, including but not limited to: reporting issues, suggesting improvements to the dataset card, sharing code for using the dataset, or contributing to the development of models trained on this dataset.