A local research suite for generating, reviewing, and training radiological report models.
Inference · Studio · Configuration · Documentation
XREPORT contains two separate applications:
| Application | Purpose | Distribution |
|---|---|---|
| XREPORT Inference | Generate and review draft reports with public models or imported checkpoints. | Windows desktop package, or from source |
| XREPORT Studio | Prepare datasets, then train and evaluate models. Exports checkpoints for Inference. | From source |
The applications have separate environments and data folders. Inference works without Studio; you need Studio only to prepare data or train your own model. Only Inference is packaged and released: the Windows packages on GitHub Releases contain Inference alone, and installing them does not install Studio. Studio is run from a repository checkout. See the changelog for what changed in each release.
Research use only. XREPORT and its models are not clinically approved. Generated text can omit findings or contain plausible errors. A qualified reviewer must check every draft against the source images and relevant context.
Use Inference to select a model, add supported images, generate a draft, and review its output. The model catalogue shows each model's input requirements, output sections, installation state, and revision. You can edit, save, copy, or export a report; saved reports retain the original output and generation metadata.
Public model weights may require a first download, and some models require provider approval and credentials. Anatomy, image limits, hardware needs, and access terms vary; consult the catalogue and model documentation. Metrics and successful generation do not establish clinical accuracy.
On Windows, download a CPU or CUDA portable executable or MSI from GitHub Releases, when available. The CUDA package requires a supported NVIDIA GPU and driver; the portable executable requires Microsoft WebView2.
To run from source, clone the repository and run this command in PowerShell from the repository root:
powershell -ExecutionPolicy Bypass -File .\start_on_windows.ps1 -Action LaunchThe launcher prepares dependencies, initializes the database, builds the frontend when needed, and opens Inference at http://127.0.0.1:5003 by default. First launch can take several minutes. Keep the launcher open while using the app; press Ctrl+C to stop it.
- Select a model and check its anatomy, image limit, output sections, and access requirements.
- Add authorized, de-identified images and choose a generation profile. The app can download an uncached public model on the first generation.
- Generate the draft, review the text and metadata, and correct or discard unsupported statements. Save edits before leaving the page.
Use Studio to build and evaluate models. Import an image folder and a report table, review matched and unmatched records, and process the dataset for training. You can start or resume training, monitor jobs, and evaluate checkpoints against their associated processed datasets.
The Studio screenshots show the Dataset and Training pages with a 12-record synthetic test dataset and a two-epoch compact checkpoint; the training dashboard is idle. The Inference screenshot is a real CheXOne run on a CC0 chest radiograph and demonstrates the interface only, not model accuracy.
_chest_radiograph_(X-ray).jpg){kind=link}
Studio runs from source. From the repository root on Windows, run:
powershell -ExecutionPolicy Bypass -File .\start_on_windows.ps1 -Action LaunchStudioStudio has its own Python environment and data folder. Its first launch installs the selected CPU or CUDA profile; an unattended first launch defaults to CPU. Keep the launcher open while using Studio and press Ctrl+C to stop it.
See Getting started for profile selection and the manual macOS/Linux development setup. The core workflows cover dataset preparation and training in more detail.
In Studio, choose Export for XREPORT on the Training page to create an .xreportckpt archive. In Inference, choose Import checkpoint in the model catalogue. This transfers the model checkpoint, not the Studio dataset or the applications' other data. See checkpoint exchange for details.
The Windows launcher creates settings/.env from the example if it is absent. The default source-mode resource directory is data; XREPORT_RESOURCES_DIR changes its location. Set HF_TOKEN only when a gated Hugging Face model requires credentials. Model downloads need network access and disk space. Back up datasets and checkpoints before removing application data.
See runtime configuration for the full set of options and troubleshooting for startup and database issues.
The source application uses Python and FastAPI for the backend, Angular for the frontend, and SQLite by default. Windows desktop releases use Tauri. Python 3.14 or newer is required for the backend; the Windows launcher prepares its own runtimes.
The project documentation index links to architecture, startup, testing, and operational guides. The Windows test runner is app\tests\run_tests.bat; its options and manual commands are described in startup and testing.
Issues and pull requests are welcome. Describe the problem and keep changes focused; the testing and quality guide covers the project's conventions.