amirmmahdavikia/pelviset-validation

PelviSet Clinical Validation Reviewer App

A mobile-friendly Gradio app for Dr. Nabian (and any future reviewers) to score the 200-case validation sample (100 random + 100 high-risk) for segmentation and keypoint plausibility. Progress saves after every case and resumes automatically, so a reviewer can do a few cases at a time on their phone and pick up where they left off.

1. Prepare the data

The app expects:

data/
  cases.csv
  cases/
    ustc_005678/
      01_raw.jpg
      02_seg.png
      03_kp.png
    ustc_005388/
      01_raw.png
      02_seg.png
      03_kp.png
    ...

This matches your existing cases/ folder as-is — just copy it under data/cases/. File extensions can vary per case/file (the app looks for .png, .jpg, or .jpeg for each of 01_raw, 02_seg, 03_kp).

data/cases.csv is your manifest (review_order, case_id, source, stratum, risk_block, age_months, age_group, sex). Only case_id is required; the rest are shown to the reviewer as context on each case and used to sort cases in review_order.

Before deploying, run the validator to catch anything missing:

python validate_cases.py

It checks every case_id in data/cases.csv has a matching directory under data/cases/ with all three files present, and flags anything mismatched.

2. Create a private dataset repo for score persistence

HF Spaces have ephemeral disks on the free tier, so scores are pushed to a small private HF dataset repo after every save (near-zero cost, and gives you a full history of every commit as a bonus audit trail).

1. 1.On huggingface.co, create a new private dataset repo, e.g. amirmmahdavikia/pelviset-validation-scores.
2. 2.Create an HF access token with write access to that repo (Settings → Access Tokens).

3. Create the Space

1. 1.New Space → SDK: Gradio → visibility: Private (recommended, since this contains real patient radiographs).
2. 2.Push these files to the Space repo: app.py, requirements.txt, data/cases.csv, data/cases/.
3. 3.If data/cases/ is large, HF will handle it via git-lfs automatically for you on push.
4. 4.In the Space's Settings → Variables and secrets, add:
5. 5.HF_TOKEN — the write token from step 2 (mark as secret)
6. 6.SCORES_DATASET_REPO — e.g. amirmmahdavikia/pelviset-validation-scores
7. 7.REVIEWERS — e.g. nabian:choose-a-password,reviewer2:another-password (mark as secret; add one user:pass pair per reviewer)
8. 8.Restart the Space. It will prompt for username/password on load — send Dr. Nabian his username/password directly (not through the Space itself).

4. Using the app

• —Log in with the reviewer credentials you set.
• —The app jumps straight to the first unscored case.
• —Three tabs per case: Radiograph / Segmentation overlay / Keypoint overlay — tabs keep each image full-width, which works better on a phone than three side-by-side images.
• —Score both Likert scales (1–5) and optionally add a note, then Save & Next. Saving is instant and pushes to the dataset repo.
• —Skip moves on without saving (for a case they want to revisit).
• —Jump to case # lets them go straight to a specific case if needed.

5. Pulling results later

At any point, download scores.csv from the dataset repo (amirmmahdavikia/pelviset-validation-scores) — it has one row per (reviewer, case_id) with seg_score, kp_score, notes, timestamp. This maps directly onto the Likert-scale validation table in the Scientific Data draft.

Notes

• —Multiple reviewers are already supported — just add more user:pass pairs to REVIEWERS. Each reviewer's progress and scores are tracked independently by username.
• —If SCORES_DATASET_REPO / HF_TOKEN aren't set (e.g. local testing), the app falls back to a local temp file so you can try it before wiring up persistence.