Medical Image Cleaner

A simple Streamlit app that uses AI image editing models (via fal.ai) to remove burned-in text, annotations, measurement lines, and other non-anatomical markings from medical images — purely through prompting, no computer vision or masking required.

The Problem

Medical image datasets often have labels, measurement lines, patient info, arrows, and other annotations burned directly into the image pixels. When training a classifier on such data, the model can learn to associate these markings with specific diagnoses (shortcut learning / overfitting) rather than the actual anatomy.

This tool preprocesses a dataset by sending each image through a generative AI model with a cleaning prompt, producing annotation-free versions ready for training.

Screenshots

Upload & Settings	Before / After Results

Models

Tier	Model	Price	Approach
Budget	FLUX.1 Dev	~$0.025/image	Strength-based img2img — regenerates the full image guided by the prompt
Standard	FLUX.1 Kontext Pro	~$0.04/image	Instruction-based editing — understands what to remove and what to keep
Premium	FLUX.2 Pro	~$0.03+/MP	Latest FLUX editor, scales with output resolution
—	Nano Banana 2 (Google)	~$0.08/image	Gemini 3.1 Flash Image — instruction-based, high fidelity

Recommendation: Start with Standard (Kontext Pro) — it's instruction-based, so it actually understands the edit rather than regenerating the whole image. Test on a handful of images before running a full dataset.

Setup

1. Clone the repo

git clone https://github.com/your-username/ImageAICleaner.git
cd ImageAICleaner

2. Create a virtual environment and install dependencies

python3 -m venv venv
source venv/bin/activate      # Windows: venv\Scripts\activate
pip install -r requirements.txt

3. Get a fal.ai API key

Sign up at fal.ai and grab a key from your dashboard.

4. Run

streamlit run app.py

The app opens at http://localhost:8501. Enter your API key in the sidebar — it's never stored anywhere.

You can also set the key as an environment variable to skip typing it each time:

export FAL_KEY=your_key_here

Usage

1. Select a model tier in the sidebar
2. Upload your images (PNG, JPG, TIFF, BMP, WEBP)
3. Review the estimated cost
4. Edit the prompt if needed (a good default is already provided)
5. Click Clean Images — results appear one by one as they finish
6. Download all cleaned images as a ZIP

Notes

• Results are stochastic — the same image may produce slightly different outputs each run
• For the Budget (FLUX Dev) model, lower the Strength slider to allow more aggressive cleaning; higher values preserve more of the original
• Generative inpainting may alter regions that overlap with anatomically relevant features — always sanity-check results before using them for training
• Images are temporarily uploaded to fal.ai's CDN for processing and are not stored permanently
• No patient data should be sent to external APIs without ensuring it has been de-identified first

Adding More Models

The model list lives at the top of app.py in the MODELS dict. Each entry needs:

"Label — Name (~$price)": {
    "endpoint": "fal-ai/model-id",
    "description": "Shown in the sidebar.",
    "prompt_style": "instructive",   # or "descriptive"
    "image_url_key": "image_urls",   # omit if the model uses "image_url" (singular)
    "params": {
        "guidance_scale": 3.5,       # any extra API params with their defaults
    },
},