PouriaRouzrokh / RadQG

In this repository, we develop automated pipelines based on Large Language Models (LLMs) for building radiology question banks.

RadQG

RadQG is an artificial intelligence (AI)-based tool to automatically generate radiology questions based on the available figures within some user-input articles from the RadioGraphics journal. RadQG works with large language models (LLMs) and this repository contains the functionalities to utilize those models for generating questions and developing a question bank.

Developer: Pouria Rouzrokh, MD, MPH, MHPE [[e-mail]]

A brief demo of our app is available below:

How to Run

This repository is set up to mostly run with a GradIO application. Please run the demo/gradio_demo.py file to run the GUI. You can change the local port number, if needed. Alternatively, you can run the demo/notebook_demo.ipynb file to run the pipeline in a notebook environment.

• Please start by installing the required packages in the requirements.txt file.

• Enter your OpenAI API key in the configs.py file.

• Feel free to change the other parameters in the configs.py file according to your needs.

• If running the demo/notebook_demo.ipynb file, please make sure to upload your RadioGraphics HTML files in the data/html_articles folder.

• Refer to the demo/notebook_demo.ipynb file or the demo/gradio_demo.py for further instructions.

Development Logs

### 11/02/2023:

- Added the `gradio_demo.py` file to include the GUI for the pipeline.
- Optimized the prompts.
- Deleted the `while` loops in the `openai.py` file.

### 10/31/2023:

- Added the "option" key for the dictionary returned from LLMs for the MCQs.

### 10/30/2023:

- Changed the name of the prompting models for clarity.
- Changed the name of the OpenAI models for clarity.

### 10/29/2023:

- Added the `generate.py` file to include the functionalities for generating the questions.
- Added the `utils.py` file to include the utility functions.
- Added the `prompts.py` file to include the prompt templates for LLMs.
- Added he `llm/openai.py` file to include the functionalities for loading the LLMs using the OpenAI API.
- Added the `demo/notebook_demo.ipynb` file to include the demo of the pipeline in a notebook environment.

### 10/28/2023:

- Reorganized the repo to reflect the new version of RadQG which will be based on Q/A from RadioGraphics figures.
- Added the `parse_html.py` file to include functionalities for HTML parsing and extracting the text, figures, and figure captions from the HTML files.
- Archived the old version of RadQG in the `archive` folder.
- Removed the `demo` folder.

### 10/27/2023:

- Set up pre-commit in the repo.
- Added `balck` and `black-nb` to the pre-commit hooks.

### 08/10/2023:

- Changed the name of the repository to "RadQG".
- Added the `main.py` file to run the pipeline using GradIO.
- Solved a bug that prevented correct query building and cosine similiarity calculation.
- Added the GIF demo of the pipeline to the `demo` folder.
- Added the `demo.mp4` and `demo.gif` files to the `demos` folder.

### 08/09/2023:

- Added the `configs.py` file.
- Added the `toy data`.
- Added the `chain_prompts.py`
- Added the `general_utils.py`
- Added the `rag.py`
- Added the `demo.ipynb`.
- Tested the pipeline on the toy data and in the demo notebook.
- Tested the functionality of the pipeline with "GPT-4" model.

### 08/08/2023:

- Created the repository.
- Cleaned and added the code for the text preprocessing.
- Added the `requirement.txt` baseline file.