1Chest Radiograph
Upload a frontal Chest X-Ray (CXR) for automated pneumonia screening.
2Assessment
Upload or select a sample Chest X-Ray on the left and click "Analyze Radiograph" to generate the diagnostic assessment and clinical report.
Ask about lungs and pneumonia
General educational questions only. This is separate from the classifier and never sees your radiograph.
Try asking:
Clinician review workspace
A small group of named clinicians review cases against one shared model and correct it where it is wrong. Access is issued by hand. There is no sign-up.
Anything that needs sorting out — a lost Rad-Intel ID, a changed phone number, a case uploaded by mistake, a question about how your data is used — goes to the study administrator personally, using the contact details you were given when your account was issued.
Submitting this creates no account and grants no access. Your details go to the study administrator, who checks them against your medical registration and contacts you.
Review a case
Run a radiograph through the shared model, then record whether it was right. Reviews are queued, exported, and only then used to train a new model version — submitting one never changes the model by itself, and training is run by the study administrator.
Review queue
Every case from every site. You can correct and delete your own; the rest are read-only.
No reviews yet. Run a case to start the queue.
An export writes a versioned manifest of every case not yet exported and freezes those records, so they keep matching what gets trained on.
What happens to reviewed cases
An export is the save point: it freezes a batch of cases so they keep matching whatever gets trained on them. Training itself is run by the study administrator, on a separate machine, and never changes the checkpoint the project reports. The held-out test partition is never trained on.
So there is nothing else for you to do here. Upload and correct cases, and press Create training export when a batch is worth keeping.
The community model —
Your details
Visible to you alone. Other clinicians see only your name and institution.
Your Rad-Intel ID is one of the two things needed to log in as you. Treat it like a password and ask the administrator for a replacement if it is seen by anyone else. Anything to change here goes through them.
Who is in the study —
So you know whose cases are in the queue. Names and sites only — no contact details.
The shared model
Which version of the project's own model is actually being served, and how fine-tuned versions compare against it.
Administrator console
Release a model to the public
A training run writes a new version and changes nothing until you do this — whether it was a fine-tuning run here or a community model trained elsewhere. Releasing one puts it in the model list on the Analyse X-ray tab, beside the models the project reports, with the scores it was measured at. The model the site uses by default does not change — people choose it deliberately. Your password is asked for every time.
Where reviewed radiographs are kept
This server's disk is rented by the month and most hosts wipe it on every redeploy. A reviewed radiograph is the one thing here that cannot be made again, so each one is also copied into your Google Drive — and fetched back from there if the disk ever comes up empty.
The community model
A second model, separate from the one the project reports. Every run discards the previous community version and trains the same architecture from scratch on the public dataset together with every case clinicians have reviewed, so a version is built from all the evidence rather than from the latest batch. The checkpoint the project's results describe is never written to.
One zip: every reviewed radiograph and a manifest of the labels. Building it freezes anything still pending, the same way an export does. The public dataset is not in it — the training script downloads that itself.
Not here. A full retrain is hours of compute and this server's job is
to serve the site. Follow COMMUNITY_MODEL.md in the
repository: it opens a free GPU notebook, takes the zip, and writes a
.pt file. The held-out partition is never trained on, so
its scores can be read beside the reported models.
It is checked, scored and stored as a new numbered version. Nothing is served and nobody can choose it — it appears in Release a model to the public above, which asks for your password again.
Add a clinician
Generate a Rad-Intel ID, create the account, then give them the ID yourself. Do not send it in the same message as the link to this site.
Access requests —
Verify each applicant against their medical registration before issuing anything. Nothing here has been granted.
Clinicians
| Name | Institution | Contact | Rad-Intel ID | Last login | Status |
|---|
Open sessions —
One device per account. A session ends when that browser closes; closing one here does the same immediately, without disabling the account.
Your administrator account
Activity log —
Logins, uploads, edits, deletions, exports and training runs, newest first. This is a record of named clinicians' behaviour, so it stays with you.
| When | Who | Action | Detail | From |
|---|
What this system is
Rad-Intel classifies paediatric chest radiographs as normal or pneumonia, shows which regions drove the score, and writes a structured summary from those recorded outputs. Classification, explanation and language generation are kept separate so that every displayed sentence traces back to a number the model produced.
Measured results
Held-out 624-image test partition, single run at seed 42.
| Configuration | Acc. | AUROC | Sens. | Spec. |
|---|---|---|---|---|
| DenseNet121 | 91.5% | 0.962 | 95.6% | 84.6% |
| Swin-T | 90.1% | 0.975 | 97.4% | 77.8% |
| ResNet50 | 89.7% | 0.954 | 93.3% | 83.8% |
| Fusion without CBAM | 93.4% | 0.977 | 94.6% | 91.4% |
| Complete classifier | 92.0% | 0.974 | 95.9% | 85.5% |
Known limitations
- Paediatric images from a single institution; nothing extends to other ages, institutions or acquisition protocols without external validation.
- Labels collapse bacterial and viral pneumonia into one class.
- No dependable patient identifier, so the split cannot guarantee patient-level independence.
- Each configuration was trained once, so no result separates an architectural effect from seed variation.
- Between 21% and 48% of model response falls outside the segmented lung fields, consistent with shortcut learning on this dataset.
- Grad-CAM and LIME agree only weakly with each other; a heatmap shows model sensitivity, not pathology.
- Generated summaries read as diagnostic to a reviewing clinician in half of sampled cases despite passing schema validation.
Which model you are using
The table above describes the five configurations trained for this project, measured on the held-out partition named there.
community_hybrid, is the same fusion architecture retrained from
scratch on the public dataset together with every contributed case. Both are
scored on the same held-out partition and their numbers are shown beside their
names, but the table above does not describe them — it
describes the five configurations only. The default is unchanged: the complete
classifier, and a released model is only ever something a visitor picks
deliberately.