Reproductive medicine runs on what the eye sees through a lens — and on trust: that a count was made carefully, a grade assigned honestly, and a patient's images kept private. FertiVision Local is an open-science attempt to give clinicians and researchers measured, reproducible, fully private image analysis — running entirely on your own machine, with WHO 2021 and Gardner criteria built in, and your signature on every result.
📸 Capture → 🔍 Measure → 📋 Classify → ✍ Your signature
The full protocol-driven workflow — study sheet, calibration lock, editable draft measurements, review queue, case timeline — lives in the workspace, one click away. The landing page stays about the why; the app shows you the how.
Semen and embryo assessment still depends on human visual grading, and the literature has long documented meaningful inter-observer variability — counts and grades that differ between trained embryologists, between labs, between days. Meanwhile, the obvious fix — sharing images and building large collaborative datasets — collides with patient privacy.
Every rule is an explicit, published cut point — WHO 2021 5th-percentile limits, Gardner & Schoolcraft grading — applied identically every time. No "it looked like a 4BB today". The criteria are inspectable open source.
Because images never leave the clinician's machine, analysis can happen anywhere — including on data that could never be uploaded to a vendor cloud. Privacy here is not a constraint; it is what makes the tool usable at all.
Every number this project states is measured on held-out expert-annotated data, with a confidence interval, and the validation code is public. When a component is not good enough yet, we say so — and the system refuses to switch it on.
Concentration and motility-relevant counts from microscopy images, classified against WHO 2021 5th-percentile reference limits — oligo-/astheno-/teratozoospermia called by explicit, published cut points you can inspect.
Ultrasound follicle counts under a per-study calibration. Longitudinal progression is computed with a fixed ±15% change flag, rates per day, and guarded math — suitable for monitoring a patient over weeks.
Expansion, ICM, and trophectoderm grades mapped to transfer recommendations per Gardner & Schoolcraft (1999), including low-confidence handling for incomplete grades.
Every photo, video, document, or note for a case becomes a timeline point under the same protocol — so the follow-up visit has the full history in one place, each point signed off by the responsible clinician.
Patient images are analysed inside your browser, on your machine. There is no upload code path at all — this is verifiable in the open source and enforced in the app. The only thing that crosses to the local server (which also runs on your machine) is structured numbers and text: counts, grades, dates, and your signature. Nothing is sent to any cloud, including ours.
| Item | Status |
|---|---|
| AI sperm detector accuracy (56% measured, 85% required) | Open — classical auditable method is used meanwhile |
| Prospective inter-rater study (app vs ≥2 blinded embryologists) | Not started — required before any clinical use |
| Regulatory clearance (this is not a medical device) | Not started |
| Report narrative drafting (needs a local install) | Works locally only; this portal is informational |
Two commands and a browser — no account, no data leaving your network. Full instructions (including Docker) are in the README:
# 1. Start the local review server
pip install -r requirements.txt
uvicorn server.app:app --port 8000
# 2. Open the app
python3 -m http.server 9000 --directory static
# → http://localhost:9000 (images stay in the browser)| Claim | Evidence |
|---|---|
| Expert-annotated training data | VISEM-Tracking (Scientific Data, 2023): 20 videos, 29,196 frames, verified in three biology labs — CC BY 4.0 |
| Semen classification criteria | WHO 2021 Laboratory Manual, 6th ed., 5th-percentile reference limits |
| Blastocyst grading criteria | Gardner & Schoolcraft (1999) |
| No image ever uploaded | Open source — the browser app contains no image-upload code; enforced by design and by test |
| Every number on this page | Reproduced by the validation suite; see full provenance |