DFL-KamLab/HSTLI_A-Dataset-of-Human-Semen-Time-Lapse-Images

HSTLI: A Dataset of Human Semen Time Lapse Images

Dataset Details

HSTLI contains 3,266 time-lapse microscopy videos of human sperm. Clips were recorded from two imaging modalities:

• —CASA system (Sperm Class Analyzer)
• —Optical microscope (Swift M10DB-MP + Fujifilm X-T30)

A subset of videos was manually annotated with bounding boxes around each visible sperm head. The dataset supports detection, tracking and motility computation.

Total contents:

• —34 annotated videos
• —~1.49 million sperm annotations
• —3,232 unannotated videos
• —~27 hours of imagery

Citation

If you use this dataset in your research, please cite the accompanying bioRxiv preprint:

HSTLI: A Dataset of Human Semen Time-Lapse Images for Detection, Tracking, and Motility Parameters Analysis Atilla Sivri, JiWon Choi, Justin Bopp, Albert Anouna, Matthew VerMilyea, Gustave Alkhoury, Omer Onder Hocaoglu, Moshe Kam, Ludvik Alkhoury bioRxiv, 2025 DOI: https://doi.org/10.64898/2025.12.15.694470

BibTeX

@article {Sivri2025.12.15.694470,
	author = {Sivri, Atilla and Choi, JiWon and Bopp, Justin and Anouna, Albert and VerMilyea, Matthew and Alkhoury, Gustave and Hocaoglu, Omer Onder and Kam, Moshe and Alkhoury, Ludvik},
	title = {HSTLI, a Dataset of Human Semen Time-lapse Images for Detection, Tracking, and Motility Parameters Analysis},
	elocation-id = {2025.12.15.694470},
	year = {2025},
	doi = {10.64898/2025.12.15.694470},
	publisher = {Cold Spring Harbor Laboratory},
	URL = {https://www.biorxiv.org/content/early/2025/12/17/2025.12.15.694470},
    abstract = {We present HSTLI, a dataset of Human Semen Time-Lapse Images acquired using two microscopy systems. These were (i) a commercial Computer-Assisted Semen Analysis (CASA) system and, (ii) an optical microscope. The dataset contains samples from 51 healthy male participants and includes 3,266 video clips (approximately 27 hours of imagery). A subset of the clips from both systems was manually annotated with bounding boxes around each visible sperm head, thereby establishing ground truth for detection, tracking, and motility analysis. Specifically, 14 CASA clips and 20 optical-microscope clips (each consisting of 900 frames) were labeled, yielding 29,950 annotated frames and roughly 1.4 million sperm annotations. The dataset also includes thousands of unlabeled clips captured under varying preparation conditions (washed vs. unwashed), dilution levels, magnifications, and regions of interest. When available, motility parameter reports are provided from either clinical technicians or CASA outputs. We demonstrate the use of this dataset in two example applications, (i) automated sperm detection and tracking using YOLOv5 and SORT, and (ii) visualizations of motility parameters. HSTLI offers a comprehensive benchmark for developing and evaluating algorithms for sperm detection, tracking, classification, and motility assessment.Competing Interest StatementL.A. holds equity in and serves as a founding engineer for Cognitive Signals Inc. These interests have been disclosed to Weill Cornell Medicine and are managed in accordance with its conflict-of-interest policies. The remaining authors declare no competing interests.},
	eprint = {https://www.biorxiv.org/content/early/2025/12/17/2025.12.15.694470.full.pdf},
	journal = {bioRxiv}
}

Contributors

• —Curated by: Atilla Sivri, JiWon Choi, Justin Bopp, Albert Anouna, Matthew VerMilyea, Gustave Alkhoury, Omer Onder Hocaoglu, Moshe Kam, Ludvik Alkhoury

Links

• —Repository: https://huggingface.co/datasets/DFL-KamLab/HSTLI_A-Dataset-of-Human-Semen-Time-Lapse-Images
• —Paper: HSTLI: A Dataset of Human Semen Time-Lapse Images for Detection, Tracking, and Motility Parameter Analysis
• —Code repository: https://github.com/DFL-KamLab/HSTLI-A-Dataset-of-Human-Semen-Time-Lapse-Images

Uses

Direct Use

Suitable use cases include:

• —Training and evaluating object detection models for sperm heads
• —Multi-object tracking in time-lapse microscopy
• —Motility parameter computation (VSL, VAP, VCL, ALH, BCF, etc.)
• —Semi-supervised or self-supervised learning based on large unlabeled videos

Dataset Structure

Files follow a BIDS-style naming convention:

sys-casa_sub-HC003_ses-01_run-001_video.avi
sys-opt_sub-HC003_ses-01_run-025_video.avi

Each video is paired with a JSON metadata file containing:

• —imaging system (CASA or optical microscope)
• —magnification level (e.g., 200x, 400x)
• —sample preparation (washed, unwashed, dilution level)
• —subject, session, and run identifiers (e.g., HC003, ses-01, run-001)
• —experiment date
• —frame rate / sampling frequency
• —scene identifier
• —device, model, and recording site information

Dataset Creation

Curation Rationale

The dataset was created to provide a benchmark for:

• —Sperm head detection and tracking
• —Comparing CASA system outputs with machine learning models
• —Studying motion patterns and motility across imaging modalities
• —Filling gaps in publicly available fertility datasets

Source Data

Raw microscopy videos were acquired at:

• —Ovation Fertility Laboratory (TN)
• —Sperm and Embryo Bank of New Jersey (SEBNJ)

Healthy adult donors provided semen samples under IRB-approved protocols.

Data Collection and Processing

• —Two imaging systems captured video at 30-60 fps
• —Clips were segmented into 30 second runs
• —Metadata was recorded for each acquisition
• —A subset was manually annotated

Who are the source data producers?

• —Adult male participants, ages 18-35
• —All samples obtained under informed consent

No personal identity or clinical history was collected.

Annotations

Annotation process

• —Annotations were created using:
• —MATLAB tooling (CASA videos)
• —Labelbox (optical microscope videos)
• —34 fully annotated videos contain ~1.49 million bounding boxes
• —Class index is always 0 (sperm)

Ground truth ships as one .txt file per frame, in a single folder per run: <run>_gt/<run>_gt_frame-<N>.txt. Each line is one bounding box:

<class> <cx> <cy> <w> <h> <track_id>

• —cx, cy, w, h are normalized to [0, 1] relative to frame width and height.
• —<class> is always 0 (sperm).
• —Frame indices <N> are 0-based.
• —<track_id> is a persistent identity: the same id recurring across frames marks the same sperm cell.

Detection and trajectory ground truth therefore live in the same file rather than in two parallel folders. To use the data purely for detection, drop the trailing field and each row is a standard YOLO row:

cls, cx, cy, w, h = line.split()[:5]      # detection
track_id          = line.split()[5]       # identity

Note for users of earlier revisions: these files previously carried five columns without the trailing track_id. The first five columns are unchanged, so readers that slice the first five fields keep working; readers that assume exactly five fields per row need the slice above.

Ground-truth trajectories

Bounding boxes are provided for all 34 manually annotated runs. Trajectory identities, which link each bounding box to the same sperm cell across frames, are carried in the track_id column of the same ground-truth files. They are provided for all 20 optical microscope runs and for 8 of the 14 CASA runs; the remainder are being added. One run is manually annotated per subject, so each row below corresponds to a single run.

Every annotated run covers 900 frames, with a single exception: sys-opt / sub-HC001 / run-016, which is annotated for its first 200 frames.

Personal and Sensitive Information

• —Dataset contains no personal identifiers
• —Images contain microscopy views only
• —All IRB and consent requirements were satisfied

Dataset Card Contact

kam@njit.edu lua4006@med.cornell.edu as3759@njit.edu