Project development is ongoing

Your CXR project,
in one workspace.

Track the dataset, saved model checkpoint, evaluation status, notebooks, and development workflow while the research continues.

Continue project ↓GitHub workflow runs ↗

Interactive research demo

Review a case

Choose an X-ray and enter the doctor’s note. The connected local model drafts from the image alone, then this page highlights report differences and lets you ask case-specific questions.

Your local model service

Not checked

Start Docker on your computer first. Your image and notes are sent to that local service and are not stored by this website.

Research only. The comparison is a limited keyword check, not a second diagnostic model; agreement does not prove correctness. A qualified clinician must review the X-ray and reports.

Current project state

What exists now

DATASET6,685 images · 3,440 studies

Study-level split has no UID overlap. Image files remain local and are not stored on this website.

Train 5,359 rowsValidation 659 rowsTest 667 rows

QWEN QLoRASaved checkpoint: step 920

About 230 optimizer updates at gradient accumulation 4. A saved step count does not establish convergence or quality.

Open training notebook ↗

EVALUATIONFresh checkpoint comparison pending

The displayed 24-case scores are legacy results from the earlier step-400 run, not step 920. They are text-overlap scores, not medical accuracy.

ROUGE-L 0.188 → 0.331BLEU 0.027 → 0.070

DEPLOYMENTProject site live · model API local

This dashboard is public. Report generation is available only from the local Docker API when it is running; this page does not upload X-rays.

Open local inference app ↗

Continue from here

Project checklist

Check off work as you complete it. These selections are saved only in this browser.

Confirm local dataset pathSet CXR_DATASET_ROOT to the existing Indiana dataset folder.Setup guide ↗ Run validation on the current checkpointUse validation studies to compare checkpoints, not the previously viewed test examples.Notebook 09 ↗ Continue QLoRA trainingResume from saved adapter/optimizer state only after data, CUDA, and disk checks pass.Training steps ↗ Run the final paired test oncePreviously exposed studies are excluded by the maintained evaluation cell.Evaluation notes ↗ Complete qualified clinical reviewReview unsupported claims, omissions, laterality, and clinically important errors.Limitations ↗

Project workflow

Open a notebook or project file

Each notebook is in the repository’s main branch. Open in VS Code/Jupyter and use the matching local environment.

00BaselinesControls and reference metrics 01Prepare dataMetadata and study splits 02CNN + attention LSTMReport-generation baseline 03Vision encoder-decoderImage-to-text baseline 04Qwen results reviewValidation and saved outputs 05Text retrievalIndication-based control 06Fusion classifierImage plus indication 07Discrepancy + Grad-CAMExploratory review 08RAG assistantTraining-report retrieval 09Qwen3-VLTraining and paired evaluation

Build and delivery

GitHub activity

Quality checks, Pages deployment, and container publishing run from the repository.

Loading recent GitHub workflow runs…

All workflow runs ↗Release v0.1.0 ↗Full README ↗

Local model app

Use the API while training continues

The Docker app runs on the project computer and loads the saved adapter. Start it from the project folder with docker compose up -d, then open the local app. The public Pages site is not connected to a hosted model.

Open local app ↗