Team T.O.P — Technology Of Prognosis
PM Kim Min-ji · CM Ji Seung-hyun · QA Seo Hyeon-jun · ENG1 Kim Tae-kyoung · ENG2 Kim Hye-hyeon
Faculty Advisor Prof. Song Gi-won | 2026-1 Convergence Capstone Design I Final Presentation | Ver 1.0.0
🌐 Language: 한국어 | English
📑 This README mirrors the structure of the final presentation deck (
deliverables/[T.O.P]Seed_최종발표_v1.0.0.pptx). It follows the table of contents and flow of the presentation slides so the whole project can be understood at a glance.
| Chapter | Contents | |
|---|---|---|
| 01 | Introduction | Company · Org chart · Project overview · Methodology |
| 02 | Problem | Problem awareness · Problem definition · Solution |
| 03 | How | Actor definition · Dev environment · System Architecture · Use Case |
| 04 | Outcome | Dataset · Model flow · Model performance · UI/UX · Demo · Impact |
| 05 | Artifacts | MS Project · Artifact status · Roadmap · Traceability · GitHub |
| 06 | Reference | References |
| + | Appendix | Kinematic metrics · WHO criteria · Model details · Datasets · Limitations |
| + | Technology Q&A | Condensed rationale for key technology choices |
| + | Quick Start | How to run · Project structure |
T.O.P stands for "Technology Of Prognosis" — carrying the meaning of predictive technology.
| Role | Name | Responsibilities | Development part |
|---|---|---|---|
| PM | Kim Min-ji (김민지) | Project oversight · Schedule · Progress management | Motility analysis model |
| CM | Ji Seung-hyun (지승현) | Dev-environment standardization · Artifact configuration · Source integration | Data collection & preprocessing |
| QA | Seo Hyeon-jun (서현준) | Artifact quality · System risk management · Model performance monitoring | Object detection model |
| ENG1 | Kim Tae-kyoung (김태경) | System architecture design/build · Integration · Performance improvement | Morphology analysis model |
| ENG2 | Kim Hye-hyeon (김혜현) | System architecture design/build · Integration · Performance improvement | Web page implementation |
SEED stands for "Sperm Evaluation and Embryo Development."
The system automatically detects and tracks sperm in a video and quantitatively evaluates their state.
Proposal → Analysis → Design → Implementation → Testing → Completion
At each stage, artifacts are produced and pass a verification step before proceeding.
- About 50% of all infertility causes are male or combined factors — infertility is not solely a female issue.
- In 2024, domestic male infertility diagnoses surpassed 100,000, rising faster than for women.
- Change in domestic infertility diagnoses (2020 → 2024): Men +36.9% · Women +28.5%
An examiner observes each sperm in the microscopy video manually and records its state.
| # | Limitation | Detail | Result |
|---|---|---|---|
| 1 | Examiner subjectivity | Results vary with skill/condition; the same sample can yield different results | → hard to achieve consistent evaluation |
| 2 | Long analysis time | Manually observing hundreds~thousands of sperm takes 30 min~1 hr | → low examination efficiency |
| 3 | Quantification limits | Motility/trajectory hard to quantify; morphology relies on subjective criteria | → hard to secure objective metrics |
CASA (Computer-Assisted Sperm Analysis) — a system that automates semen analysis using microscopy video and image-analysis algorithms.
| # | Limitation | Detail | Result |
|---|---|---|---|
| 1 | Morphology limits | Morphology reading still relies on manual, subjective expert judgment | → time cost and consistency issues |
| 2 | Separate morphology/motion | Morphology and motility analyzed separately, no per-same-sperm result | → integrated evaluation is hard |
| 3 | Low accessibility | Expensive equipment (USD 30–40k) · specialists needed | → barriers to adoption in general clinics |
| Product type | Representative | Key limitation |
|---|---|---|
| Domestic instant kit | TENGA Men's Loupe | Mostly sperm concentration · cannot measure morphology/volume/pH |
| Overseas instant kit | YO Home Sperm Test (FDA-cleared) | Cannot measure the key metric, sperm morphology |
| Mail-in kit | Fellow Semen Analysis | Wait time for results · morphology and other key parameters missing |
| # | Problem | Detail | Result |
|---|---|---|---|
| 1 | Lack of result objectivity | Manual-reading variance; CASA morphology depends on expert subjectivity | → time cost and consistency issues |
| 2 | No integrated morphology/motility analysis | Separate analysis; no per-sperm-object integrated analysis | → fragmented analysis limits holistic evaluation |
| 3 | Low accessibility | Expensive equipment/specialists required | → constraints for general clinics and individuals |
| # | Contribution | Core technology | Target performance |
|---|---|---|---|
| 1 | Quantified motility analysis | YOLO11 detection + ByteTrack tracking + VCL·VSL etc. kinematics* | mAP@50 ≥ 0.65 · MAE ≤ 7.0 |
| 2 | Integrated morphology/motility | EfficientNet-B3 morphology + motility·morphology module integration | Per-part mean AUC ≥ 0.72 |
| 3 | Improved accessibility | Works from a standard microscopy video · web-based results without costly equipment | — |
*Kinematics: characteristics of movement — quantifying how fast, along what path, how straight, and how widely a sperm moves.
| Actor | Situation | Goal |
|---|---|---|
| Male user | A patient planning pregnancy who wants to check infertility in advance | Self-check without dedicated equipment/personnel |
| Medical staff | Clinicians needing a reference aid for semen analysis | Objective, quantitative numbers to assist reading |
| Category | Technology |
|---|---|
| Language | Python 3.10 |
| Dev tools | JupyterLab · Cursor IDE |
| AI framework | PyTorch · Ultralytics (YOLO11) · scikit-learn |
| Web · API server | Flask · Gunicorn |
| External deploy | ngrok · Render |
| Version control | Git · GitHub |
Data-flow view — module mapping:
┌─────────────────────────────────────────────────────────────────┐
│ [Input layer] microscopy video (.mp4/.avi) → quality · normalize │
│ quality.py · normalizer.py → 640×480 / 50fps │
├─────────────────────────────────────────────────────────────────┤
│ [Detect·track] YOLO11 detection → ByteTrack ID·trajectory │
│ detector.py · tracker.py │
├─────────────────────────────────────────────────────────────────┤
│ [Analysis] kinematics · motility regression · morphology │
│ casa_features.py · analyzer.py · morphology.py │
├─────────────────────────────────────────────────────────────────┤
│ [Interpret·out] WHO verdict · confidence · annotated video · report │
│ interpreter.py · annotator.py · webapp/ │
└─────────────────────────────────────────────────────────────────┘
When a user uploads a microscopy video on the web, the server runs the analysis asynchronously and returns a result screen with motility, morphology, and kinematic metrics plus a confidence score.
Total raw scale: 105 videos (VISEM-Tracking 20 · VISEM 85), ~29,000 annotated frames, 1,540 morphology crop images.
| Purpose | Dataset | Provider | Data type |
|---|---|---|---|
| Detection·tracking | VISEM-Tracking | SimulaMet · OsloMet (2023) | Microscopy video (640×480, 50fps) + bounding boxes·tracking IDs |
| Motility analysis | VISEM (original) | SimulaMet · OsloMet (2019) | Microscopy video + motility measurements (CSV), 85 participants |
| Morphology analysis | MHSMA | Javadi & Mirroshandel (2019) | Sperm crop images (128×128) 1,540 + labels for 4 parts |
- VISEM family — published in Nature Scientific Data · CC BY 4.0. Adopted as an international standard (MediaEval) → comparable under the same conditions as world-leading research.
- MHSMA — a standard benchmark for sperm morphology analysis. Its unstained microscopy condition matches real analysis videos.
| # | Stage | Model / Algorithm | Role |
|---|---|---|---|
| ① | Object detection | YOLO11 | Per-frame sperm detection |
| ② | Object tracking | ByteTrack | Restore same-sperm IDs·trajectories |
| ③ | Kinematics | CASA metric computation | Kinematic metrics (VCL/VSL/ALH etc.) |
| ④ | Motility analysis | Ridge + RandomForest ensemble | Predict progressive·non-progressive·immotile ratios |
| ⑤ | Morphology analysis | EfficientNet-B3 | Classify head·acrosome·vacuole·tail (4 parts) |
| Item | Metric | Result | Target | Meaning |
|---|---|---|---|---|
| Motility | MAE | 6.90 %p | ≤ 7.0 | Mean difference between predicted motility (%) and measured value |
| Detection | mAP@50 | 0.677 | ≥ 0.65 | Rate of correctly detecting sperm at the right location |
| Morphology | Mean AUC | 0.727 | ≥ 0.72 | Ability to distinguish normal/abnormal per sperm part |
- Motility MAE 6.90 %p — a 0.41 %p reduction vs. the prior best under the same conditions, motilitAI (7.31 %p) (5-Fold CV).
- Both the detection and morphology models exceeded their targets.
📈 For per-version comparison, cross-validation, and domain-adaptation experiments, see
docs/performance.md.
① Upload screen → ② Analysis screen (async progress) → ③ Result screen (metrics·annotated video)
Implemented as a Flask-based web application (webapp/ · app.py).
The full analysis process from video upload to result delivery is demonstrated in the web demo.
▶ Click the image to play the demo video (
docs/images/demo.mp4). (See Quick Start below for how to run the demo.)
| # | Effect | Detail |
|---|---|---|
| 1 | Objectivity | Consistent results independent of examiner skill/condition · objective, quantitative reading aid |
| 2 | Integrated analysis | Simultaneous per-same-sperm morphology·motility evaluation → holistic evaluation previously impossible |
| 3 | Accessibility | Analysis from a standard microscopy video without costly equipment → extends even to self-check settings |
Waterfall stage schedules were managed with MS Project, completed at 100% progress.
Original file: deliverables/[PM]T.O.P_Ms_Project_v1.0.0.mpp
| Stage | Artifact | Author |
|---|---|---|
| Proposal | Team formation · Logo selection · Project proposal | PM Kim Min-ji |
| Artifact management plan | CM Ji Seung-hyun | |
| Analysis | Configuration management plan | CM Ji Seung-hyun |
| Risk management plan | ENG1·ENG2 | |
| Quality management plan · Requirements review | QA Seo Hyeon-jun | |
| Initial project development plan | PM Kim Min-ji | |
| Requirements analysis | ENG1·ENG2 | |
| Design | Basic design · Detailed design | ENG1·ENG2 |
| Implementation | Source-code documentation | CM Ji Seung-hyun |
| Testing | Unit test plan·result · Integration test plan·result | QA Seo Hyeon-jun |
| Completion | Development completion report | PM Kim Min-ji |
| User manual | ENG1·ENG2 | |
| Other | Final artifact template | CM Ji Seung-hyun |
| MS Project | PM Kim Min-ji |
📄 The final integrated artifacts are preserved in
deliverables/[CM]T.O.P_통합산출물_v1.0.0.hwp(PDF included).
| # | Direction | Detail |
|---|---|---|
| 1 | Faster processing | Parallelize video preprocessing · real-time processing → shorter wait time |
| 2 | Low-concentration handling | Compensate accuracy drop when sperm count < 10 · refine confidence penalties · improve re-capture guidance |
| 3 | Parameter expansion | Add currently-unmeasured items (concentration·total count·vitality) → higher WHO 6th Ed. completeness |
The whole project's schedule, artifacts, and issues were shared and tracked via MS Teams, securing configuration management and collaboration traceability.
Source code and system configuration management: https://github.com/MoriochoRadio/seed-project
View all references (20)
- Leslie, S. W., et al. (2024). Male infertility. StatPearls.
- National Health Insurance Service. (2024). Infertility diagnoses over the past 5 years. National audit data.
- Liang, Y., et al. (2025). Global, regional, and national prevalence and trends of infertility (1990–2021). Human Reproduction, 40(3), 529–544.
- Bieniek, J. M., et al. (2021). A Novel Approach to Improving the Reliability of Manual Semen Analysis. PMC.
- Siddharth, K., et al. (2023). Interobserver Variability in Semen Analysis. Cureus, 15(10), e46388.
- Barcena, P., et al. (2025). AI approaches to male infertility in IVF: a mapping review. PMC.
- Chawre, S., et al. (2024). A Review of Semen Analysis: Updates From the WHO Sixth Edition Manual. Cureus, 16(6), e63485.
- Mortimer, D., & Mortimer, S. T. (2015). The future of computer-aided sperm analysis. Asian Journal of Andrology, 17(4), 545–553.
- Finelli, R., et al. (2021). The validity and reliability of computer-aided semen analyzers: a systematic review. Translational Andrology and Urology, 10(7), 3069–3079.
- Gonzalez, D., et al. (2021). Clinical Update on Home Testing for Male Fertility. World Journal of Men's Health, 39(4), 615–625.
- Kobori, Y., et al. (2016). Novel device for male infertility screening with single-ball lens microscope and smartphone. Fertility and Sterility, 106(3), 574–578.
- U.S. FDA. (2016). K161493: YO Home Sperm Test (510(k)).
- Samplaski, M. K., et al. (2021). Development and validation of a novel mail-in semen analysis system. Fertility and Sterility, 115(4), 922–929.
- Thambawita, V., et al. (2023). VISEM-Tracking, a human spermatozoa tracking dataset. Data in Brief, 47, 108944.
- Hicks, S. A., et al. (2019). Machine learning-based analysis of sperm videos and participant data for male fertility prediction. Scientific Reports, 9, 16770.
- Javadi, S., & Mirroshandel, S. A. (2019). A novel deep learning method for automatic assessment of human sperm morphology. Computers in Biology and Medicine, 109, 182–194.
- World Health Organization. (2021). WHO laboratory manual for the examination and processing of human semen (6th ed.).
- Zhang, Y., et al. (2022). ByteTrack: Multi-Object Tracking by Associating Every Detection Box. ECCV 2022.
- Tan, M., & Le, Q. V. (2019). EfficientNet: Rethinking Model Scaling for CNNs. ICML 2019, PMLR 97, 6105–6114.
- Miahi, E., et al. (2019). Genetic Neural Architecture Search for automatic assessment of human sperm images. arXiv:1909.09432.
📐 Kinematic metrics ① — Velocity (how fast and straight a sperm moves)
| Metric | Name | Definition | Computation |
|---|---|---|---|
| VCL | Curvilinear velocity | Speed along the actual winding path | actual path length ÷ time |
| VSL | Straight-line velocity | Speed based on start → end straight distance | straight displacement ÷ time |
| VAP | Average-path velocity | Speed along a smoothed average path | smoothed path ÷ time |
Computed right after tracking (ByteTrack) · unit µm/s · based on 50 fps
📐 Kinematic metrics ② — Linearity·oscillation (what pattern it moves in)
| Metric | Name | Meaning | Computation |
|---|---|---|---|
| LIN | Linearity | How straight the path is | VSL ÷ VCL |
| STR | Straightness | Straight ratio vs. average path | VSL ÷ VAP |
| WOB | Wobble | Average path vs. actual path | VAP ÷ VCL |
| ALH | Amplitude of lateral head displacement | Side-to-side swing width (µm) | lateral deviation vs. average path |
| BCF | Beat-cross frequency | Tail crosses per second (Hz) | lateral-signal zero crossings ÷ time |
| PAW | Mean lateral width | Mean side-to-side path width (µm) | mean of lateral deviation |
🏃 Motility grades & WHO criteria
| Grade | Name | Description |
|---|---|---|
| PR | Progressive motility | Moves straight forward well — most important for fertilization |
| NP | Non-progressive motility | Moves but inefficiently (in place, circular, etc.) |
| IM | Immotile | Barely moves |
WHO normal criteria — total motility (PR+NP) ≥ 40% · progressive (PR) ≥ 32%
🤖 The 4 AI models & training
| # | Role | Model | Description | Training |
|---|---|---|---|---|
| 1 | Object detection | YOLO11 | Quickly finds and boxes small sperm | VISEM-Tracking videos split into frames → supervised box learning |
| 2 | Object tracking | ByteTrack | Same ID across frames → trajectory restoration | No separate training (rule-based box linking) |
| 3 | Motility analysis | Ridge + RF ensemble | Predicts motility % from trajectory features; averaged for stability | Kinematic features of trajectories → regression on VISEM clinical motility (CSV) |
| 4 | Morphology analysis | EfficientNet-B3 | Judges 4-part normality from the sperm image | MHSMA 1,540 crops, 4-part classification + VisemStyleAugment augmentation |
📏 Meaning of the metrics (mAP · MAE · AUC)
| Metric | Meaning | Direction | This system |
|---|---|---|---|
| mAP@50 | How accurately sperm are found (detection accuracy) | higher is better ↑ | detection 0.677 |
| MAE | Mean difference between predicted motility% and actual | lower is better ↓ | motility 6.9 %p |
| AUC | Normal/abnormal discrimination (0.5 = chance, 1.0 = perfect) | higher is better ↑ | morphology 0.727 |
🔬 Morphology — the 4 sperm parts
| Part | English | Description |
|---|---|---|
| 머리 | head | Shape/size normality |
| 첨체 | acrosome | Enzyme sac at the front of the head — needed to penetrate the egg |
| 공포 | vacuole | Empty space inside the head — a defect if present |
| 꼬리 | tail | The part responsible for movement |
Normal morphology rate — proportion of sperm normal in all 4 parts · WHO ≥ 4% normal · reference-only due to domain gap
📊 Confidence score · Video normalization · 5-second analysis
Confidence score (0–100) — how much the analysis result can be trusted
- Purpose: quantify result reliability on a 0–100 scale
- Penalties: few sperm (N<10) · unstable detection · short tracking
- Use: below 60 outputs a 're-capture recommendation'
- Example: a video with only 2 detected sperm → confidence 50 → re-capture recommended
Video normalization — automatically converts videos of various resolutions/frame rates to the training domain (640×480·50fps) so the model behaves consistently
5-second (250-frame) analysis — motility·CASA statistics converge stably over short spans. Clinical CASA also analyzes short fields → ensures speed·consistency
⚠️ System limitations
| Item | Limitation |
|---|---|
| Morphology | Reference-only numbers due to the training↔application domain gap (lower confidence than motility) |
| Low quality·concentration | Few sperm or blurry video can cause error → mitigated by the confidence score |
| Purpose | An assistive/reference tool; does not replace medical diagnosis |
A condensed, answer-ready summary of the technology-choice rationale scattered across the main text and the Appendix. See each link for the detailed grounds.
Q1. Why the YOLO11 + ByteTrack combination for detection and tracking? YOLO11 rapidly detects small sperm frame by frame, and ByteTrack recovers each sperm's identity and trajectory purely through detection-box association rules, with no extra training. Those trajectories are what make kinematic metrics such as VCL and VSL computable. → Appendix — 4 AI models & training
Q2. Why a Ridge + RandomForest ensemble for motility prediction? The task is regressing kinematic features from the tracked trajectories onto VISEM's clinical motility values, and averaging the two models stabilizes the prediction. The result was an MAE of 6.90 %p, lower than motilitAI (7.31 %p), the prior study under identical conditions. → Model Performance
Q3. Why choose the VISEM family and MHSMA as datasets? The VISEM family is a standard dataset published in Nature Scientific Data and adopted by MediaEval, enabling comparison with the world's best research under identical conditions; MHSMA is the standard benchmark for morphology analysis whose unstained, microscope conditions match our actual analysis videos. → Dataset Description
Q4. Why analyze only 5 seconds (250 frames) instead of the whole video? Motility and CASA statistics converge stably even over short segments, and clinical CASA likewise analyzes short fields, so this choice secures both speed and consistency. → Appendix — Confidence score · Video normalization · 5-second analysis
Q5. Why are the morphology numbers limited to "reference only"? Because of the domain gap between the training data (MHSMA) and the applied videos, morphology is less reliable than motility, so it is explicitly marked reference-only; errors on low-quality or low-concentration videos are mitigated by the confidence score (re-recording recommended below 60 points). → Appendix — System limitations
# 1. Clone the repo
git clone https://github.com/MoriochoRadio/seed-project.git
cd seed-project
# 2. Create a virtual environment and install dependencies
conda create -n seed python=3.10
conda activate seed
pip install -r requirements.txt
# Install PyTorch separately for your environment (see the comment at the top of requirements.txt)
# 3. Prepare model weights
# - Motility/morphology models are included under models/
# - YOLO11 weights (best.pt) auto-download from Google Drive
# (after setting GDRIVE_* env vars) python setup_models.py
# 4. Run the web demo
python app.py # http://localhost:5000Deployment is configured for Render.com via
render.yaml.
seed-project/
├── app.py ← Flask entry point
├── src/ ← core analysis modules
│ ├── detector.py (YOLO11 sperm detection)
│ ├── tracker.py (ByteTrack tracking + motility features)
│ ├── casa_features.py (CASA kinematic metrics)
│ ├── analyzer.py (motility ensemble regression)
│ ├── morphology.py (EfficientNet-B3 morphology analysis)
│ ├── interpreter.py (WHO-based verdict · confidence)
│ ├── normalizer.py (video normalization) · quality.py (quality assessment)
│ ├── annotator.py (annotated video generation) · pipeline.py (integrated pipeline)
│ └── ...
├── webapp/ ← Flask web application (routes · templates · static)
├── models/ ← trained model weights (motility ensemble, morphology v3)
├── notebooks/ ← experiment·training notebooks (01~16)
├── docs/ ← project docs (architecture · performance · guides)
├── deliverables/ ← final presentation materials & official artifacts (v1.0.0)
└── data/ ← datasets (gitignored)
- System design —
docs/architecture.md - Detailed performance —
docs/performance.md - User / Admin guides —
docs/USER_GUIDE.md·docs/ADMIN_GUIDE.md - Final presentation · artifacts —
deliverables/
This project is an academic capstone project and is not a clinical diagnostic tool.
The distribution license is LICENSE (MIT).
Team T.O.P · Technology Of Prognosis · 2026