RT-Rename

RT-Rename standardizes radiotherapy structure names against TG-263 using local or cloud-hosted LLMs. The repository is now organized so the web UI, parsing logic, DICOM helpers, and inference orchestration are separated, which makes it much easier to extend toward future VLM support.

What Changed

• Core logic now lives in the rt_rename/ package instead of a single utils.py module.
• The Dash UI is isolated in rt_rename/web.py.
• app.py and batch_rename.py remain as thin entry points.
• Parsing, prompt rendering, guideline loading, inference, exports, and DICOM updates are split into focused modules.
• Basic automated tests were added under tests/.

Repository Layout

rt_rename/
  config.py          Model registry loading
  constants.py       Shared paths and defaults
  dicom_utils.py     RTStruct read/write helpers
  exports.py         CSV export helpers
  guidelines.py      TG-263 workbook loading
  inference.py       Local/cloud inference adapters
  parsers.py         CSV, DICOM, and filename parsing
  prompts.py         Prompt template rendering
  rename_service.py  End-to-end rename orchestration
  web.py             Dash app layout and callbacks
app.py               Web entry point
batch_rename.py      CLI batch entry point
utils.py             Backward-compatible shim
tests/               Unit tests for core logic

Installation

Docker

1. Clone the repository.
2. Start the stack:

cd docker
docker-compose up -d

3. Open the app at http://localhost:8055.

If you want cloud inference, create a root .env file with:

OPEN_AI_URL=your_api_url_here
OPEN_AI_API_KEY=your_api_key_here

Local Development

1. Create a virtual environment.
2. Install dependencies:

pip install -r docker/requirements.txt

3. Run the web app:

python app.py

Development Workflow

Run Tests

python -m unittest discover -s tests

Batch Rename

The batch entry point is now a CLI instead of a hard-coded script.

python batch_rename.py input.csv output.csv \
  --model "Llama 3.1 | 70B | local" \
  --prompt prompt_latest.txt \
  --guideline TG263 \
  --region Thorax

For DICOM RTStruct input you can also export a renamed RTStruct:

python batch_rename.py input.dcm output.csv \
  --model "Llama 3.1 | 70B | local" \
  --prompt prompt_latest.txt \
  --output-dicom renamed_rtstruct.dcm

Configuration

Models

config/models.json defines the available models shown in the UI. The loader now supports provider-aware metadata, so we can extend the registry later for multimodal or VLM-capable models without rewriting the app architecture.

Current fields:

• name
• parameters
• model_str
• cloud

Optional future-facing fields already supported by the loader:

• provider
• modalities

Prompts

Prompt templates live in config/prompt_*.txt and are rendered by rt_rename/prompts.py.

TG-263 Nomenclature

The TG-263 workbook is stored at config/TG263_nomenclature.xlsx.

VLM Readiness

The codebase is in a better position for future VLM work because:

• inference requests are funneled through a dedicated adapter layer
• cloud requests already support structured content parts
• UI state is less dependent on module-level globals
• pure rename logic is testable without importing the Dash app

The next natural step for VLM support would be extending rt_rename/inference.py and the model registry so image-bearing requests can flow through the same orchestration path.