A Dataset of Human Sperm Microscopy Videos for Detection, Tracking, and Motility Parameter Analysis

Technical Summary: HSMV Dataset for Human Sperm Microscopy

Problem Statement
Semen analysis is a fundamental component of male fertility evaluation, traditionally performed manually by technicians using optical microscopes. Despite adherence to World Health Organization (WHO) guidelines, manual assessment is prone to subjectivity and inter-laboratory variability due to differences in operator skill. While Computer-Assisted Semen Analysis (CASA) systems aim to mitigate these issues through automated image processing, recent reviews highlight a need for further validation and standardization to improve accuracy and reproducibility. A critical bottleneck in advancing CASA and related computer vision algorithms is the lack of high-quality, diverse, and richly annotated datasets that reflect the complexity of real-world laboratory conditions. Existing open-source datasets (e.g., VISEM, SVIA, Hi-LabSpermTracking) often suffer from limitations such as a lack of annotated frames, restriction to specific clinical populations (e.g., obesity), absence of sample preparation metadata, or low frame rates.

Methodology and Data Collection
The authors present HSMV, a comprehensive dataset of Human Sperm Microscopy Videos designed to address these gaps. The data was collected from two distinct sites using two different imaging modalities:

1. Optical Microscope: Recorded at the Sperm and Embryo Bank of New Jersey (SEBNJ) using a Swift M10DB-MP Series microscope with a Fujifilm X-T10 camera.
2. CASA System: Recorded at the Ovation-Fertility Laboratory using a Sperm Class Analyzer (SCA) by MICROPTIC S.L.

Data Characteristics:

• Participants: 51 healthy male participants (24 from optical microscope, 27 from CASA).
• Volume: 3,161 video clips totaling approximately 26 hours of imagery.
• Preparation Conditions: The dataset includes samples under varying conditions: raw/unwashed, chemically washed (via density gradient centrifugation), and diluted (1:5 and 1:10 ratios).
• Imaging Parameters:
  • Optical Microscope: 100x and 200x magnification, 1920x1080 resolution, 59.94 fps.
  • CASA: 200x and 400x magnification, 1280x1024 resolution, 30 fps.
• Annotations: A subset of clips was manually annotated to establish ground truth.
  • Optical Microscope: 20 clips (one per participant, partially annotated) with 900 frames each (approx. 15 seconds).
  • CASA: 14 clips (one per participant, fully annotated) with 900 frames each (30 seconds).
  • Total Annotations: Approximately 1.49 million sperm annotations across 29,900 annotated frames (noting that one optical microscope clip was annotated for only 200 frames rather than the full 900).
  • Format: Annotations include bounding boxes (normalized YOLO format: class, center x, center y, width, height) and trajectory IDs to link cells across frames.
• Metadata: The dataset follows a BIDS-inspired directory structure. Each recording is paired with a .json metadata file containing acquisition parameters (magnification, dilution, sample type, device details) and is accompanied by motility parameter reports (volume, concentration, motility percentages) generated by either clinical technicians or the CASA system.

Key Contributions

1. Dual-Modality Dataset: HSMV is the first publicly available dataset to provide manually annotated sperm recordings from both optical microscope and CASA systems, paired with motility metadata.
2. Diversity and Realism: Unlike previous datasets, HSMV includes samples from the general population (not just specific clinical cohorts) and captures a wide range of preparation conditions (washed vs. unwashed, various dilutions, and magnifications).
3. High-Quality Ground Truth: The dataset includes rigorous manual annotations verified by multiple team members, providing a reliable benchmark for detection and tracking algorithms.
4. Reproducibility Tools: The authors provide Python scripts for running YOLOv5 and SORT (Simple Online and Realtime Tracking), as well as a toolkit (pycasa) for processing semen videos, computing motility parameters, and generating visualizations.

Results and Validation
The paper demonstrates the utility of the dataset through two example applications:

• Object Detection: A YOLOv5 model trained on the optical microscope subset achieved a precision of 0.66, recall of 0.76, and F1-score of 0.71 on a held-out test set. When applied to CASA recordings (using weights trained on optical data), the model achieved a precision of 0.82 but a lower recall of 0.43 (F1-score: 0.56), highlighting the domain shift between imaging systems.
• Tracking and Kinematics: Using the SORT algorithm on both ground truth and YOLOv5 detections, the authors computed eight standard kinematic parameters (VCL, VSL, VAP, LIN, WOB, STR, ALH, MAD). Visualizations (spider plots) show the agreement between parameters derived from manual ground truth versus automated detection within each system.

Significance and Claims
The authors position HSMV as a foundational resource for developing and evaluating next-generation CASA systems and automated semen analysis workflows. By offering a benchmark that spans multiple imaging modalities and includes robust manual annotations alongside clinical motility reports, the dataset aims to support:

• The development of more accurate, reproducible, and generalizable automated detection and tracking algorithms.
• Research into unsupervised and semi-supervised learning approaches using the large volume of unannotated data.
• Training and calibration of laboratory technicians by comparing their assessments against automated outputs and ground truth.

The paper modestly notes that the manual annotations represent ground truth based on human judgment and are subject to annotator expertise. For applications requiring exact ground truth, the authors suggest complementing HSMV with synthetic data generation tools. The dataset is intended as a "work in progress," with future releases planned to include additional recordings and annotations.