arup-ri/kinyoun_afb_50k

This dataset consists of image tiles extracted from whole-slide images (WSIs) of Kinyoun-stained smears, intended for the detection of acid-fast bacteria (AFB) such as Mycobacterium tuberculosis and related species.

The obj_det_train and obj_det_test tarballs contain approximately 50k tiles intended for training an object detection model, split into training and test sets. Roughly 20% of these tiles contain annotated AFB with bounding boxes drawn by trained AFB laboratory experts. The remainder are from verified AFB-negative WSIs and contain difficult background debris and stain artifacts to create challenging negative examples for training & testing a model.

The slide_pred_val tarballs contain approximately 1.8 million unannotated tiles sampled randomly from the central 1 cm^2 of WSIs. This is intended as a validation set for making slide-level predictions aggregated from AFB object detections. Note slide_pred_val is split into chunks because of its size (~100Gb). To reassemble them after downloading, cat slide_pred_val.tar.gz.* | tar xvfz - or similar command should work on most *nix-like systems (expect this extraction to take 15 min or more, even with fast disk io). Note you'll need over 200Gb of free space to download & then combine the chunks!

More details can be found in the paper (in press, citation TBA).

Dataset Details

Dataset Description

• Curated by: Applied AI & Bioinformatics group within the Research & Innovation unit of ARUP Labs
• Funded by: ARUP Laboratories
• License: CC BY-NC-SA 4.0

Dataset Sources

• Paper: (in press, citation TBA)
• Code: https://github.com/arup-ri/afb_detection
• Models trained on this dataset: https://huggingface.co/arup-ri/afb

Uses

Direct Use

The intended use case for this dataset, described in the paper, is training and validating a model for object detection of AFB on Kinyoun-stained WSIs.

Out-of-Scope Use

For research use only, this is not intended to support clinical use!

Dataset Structure

Image and label filenames have the following pattern:

{specimen ID}_{WSI ID}_{left}_{top}_{width}_{height}.{txt or png}

• specimen ID: refers to a unique clinical sample
• WSI ID: refers to a unique WSI. Note that in many instances multiple WSIs correspond to the same specimen ID since the same physical glass slide was scanned multiple times on different scanners and/or at different optical magnifications
• left/top: x/y coordinates of the upper-left corner of the tile in global pixel coordinates of the original WSI
• width/height: width and height of the tile in pixels ON THE ORIGINAL WSI

Note that all image tiles in the dataset are 256 x 256 pixels at a physical size of 0.2878 microns per pixel (MPP). However, many of the widths/heights in the filenames are not 256. This is because the original WSIs were scanned at a variety of different optical magnifications and MPPs, and when tiles were extracted they were resized to a consistent MPP. The left/top/width/height values can be used to reconstruct the absolute position and overlap of tiles in the original WSI they were extracted from.

Files are grouped into folders based on which WSI they were extracted from. The contents of the labels/ directory matches one-to-one with the images/ directory, with a .txt file containing bounding box annotations (if any) for the corresponding .png image.

Dataset Creation

Curation Rationale

The intended use for this dataset was to train an object detection model for screening of AFB clinical laboratory specimens, but it was decided that the model's performance was not sufficient for clinical deployment.

Source Data

Data Collection and Processing

In our standard clinical workflow, an initial screen is performed on specimens using manual microscopy on auramine O (AO)-stained slides. A random subset of AO+ specimens were selected for annotation, and separately a random subset of AO- specimens were selected for validation and hard negative training. Kinyoun-stained slides were made directly from the selected specimens (not part of the normal laboratory workflow).

After slides were digitized, tiles were extracted and resized to consistent microns per pixel using in-house python libraries. AFB-negative tiles in the object detction training and test splits were a randomly selected subset of tiles on which an early iteration of our object detection model made false-positive predictions. See the paper for more details.

Who are the source data producers?

This dataset is sourced from de-identified clinical specimens from the AFB/Mycology laboratory at ARUP Laboratories. Patient demographic data is unavailable for specimens in this dataset.

Annotations

Annotation process

AFB experts at ARUP Laboratories manually reviewed images and drew bounding box annotations using Digital Slide Archive and LabelStudio. Objects were annotated as belonging to one of 4 categories: AFB, AFB-mimic, non-AFB, or Unknown. AFB-mimic included other organisms such as Streptomyces that might look similar to AFB to a computer vision model but which should not be regarded as an AFB-positive diagnosis. Unknown includes objects for which the annotator was unsure if the true label should be AFB, AFB-mimic, or non-AFB. Note the vast majority of annotated objects in the dataset are AFB.

Who are the annotators?

Staff from the Research & Innovation unit and AFB/Mycology laboratory at ARUP Laboratories: Elizabeth Enrico, Jeffrey Gilivary, Amanda Vance, Haleina Hatch, Blaine Mathison (see paper, link TBA)

Personal and Sensitive Information

This dataset contains only image tiles extracted from WSIs (instead of entire WSIs) which are free of sensitive information. Specimen IDs found in filenames in this dataset are hashed unique identifiers, rendering it impossible to identify any information about patients.

Bias, Risks, and Limitations

Although there are >1.8M individual image tiles in this dataset, the number of distinct specimens (251) and unique patients (203) represented in the dataset is relatively small, which may cause bias and hinder its generalizability for other use cases.

Annotators reported that, compared to the standard clinical workflow using AO-stained slides, it was more challenging to confidently distinguish AFB from background debris, stain artifacts, etc., in some of these Kinyoun-stained WSIs. This was especially a problem at lower optical magnifications and/or in regions of WSIs which were mildly to moderately out-of-focus. So it is likely there is some label noise in the annotations and some number of true AFB in the images which are unannotated, though this is difficult to quantify.

Recommendations

For research use only! This dataset is not intended for clinical-grade workflows.

Citation

BibTeX:

@article{english_use_2025,
    title = {Use of a convolutional neural network for direct detection of acid-fast bacilli from clinical specimens},
    volume = {0},
    url = {https://journals.asm.org/doi/10.1128/spectrum.00602-25},
    doi = {10.1128/spectrum.00602-25},
    number = {0},
    urldate = {2025-06-25},
    journal = {Microbiology Spectrum},
    author = {English, Paul and Morrison, Muir J. and Mathison, Blaine and Enrico, Elizabeth and Shean, Ryan and O'Fallon, Brendan and Rupp, Deven and Knight, Katie and Rangel, Alexandra and Gilivary, Jeffrey and Vance, Amanda and Hatch, Haleina and Lin, Leo and Ng, David P. and Shakir, Salika M.},
    month = jun,
    year = {2025},
    note = {Publisher: American Society for Microbiology},
    pages = {e00602--25},
    file = {Full Text PDF:/Users/paul.english/Zotero/storage/PBFKYY88/English et al. - 2025 - Use of a convolutional neural network for direct detection of acid-fast bacilli from clinical specim.pdf:application/pdf},
}