🫁 CXR-Findings-AI — Chest X-Ray Findings Generator + Attention Explorer

Live Demo (CPU-Only, thanks to Hugging Face Spaces)

🔗 https://huggingface.co/spaces/manu02/CXR-Findings-AI

---

🧠 Overview

CXR-Findings-AI is an interactive Gradio application that:

✅ Generates radiology findings from a chest X-ray image

✅ Visualizes multimodal attention (image ↔ text) across layers and heads

✅ Runs entirely on CPU, showcasing the efficiency of the underlying 246M-parameter model

The system lets researchers, clinicians, and students explore how different image regions influence each generated word, enabling deeper interpretability in medical AI.

---

🔍 What This App Provides

🫁 1. Findings Generation

A lightweight multimodal model produces chest X-ray findings directly from the uploaded image.

👁️ 2. Layer-wise & Head-wise Attention Visualization

Inspect how the model distributes attention:

• Across transformer layers
• Across attention heads
• Between image tokens (32×32 grid → 1024 tokens)
• And generated text tokens

🎨 3. Three Synchronized Views

For each selected word:

1. Original Image
2. Overlay View: Image + blended attention map
3. Pure Heatmap: Visualizes raw attention intensities

🧩 4. Word-Level Interpretability

Click any word in the generated report to reveal its cross-modal attention patterns.

---

🚀 Quickstart (Local Usage)

1) Clone

git clone https://github.com/devMuniz02/Image-Attention-Visualizer
cd Image-Attention-Visualizer

2) (Optional) Create a virtual environment

Windows:

python -m venv venv
.\venv\Scripts\Activate.ps1

macOS / Linux:

python3 -m venv venv
source venv/bin/activate

3) Install requirements

pip install -r requirements.txt

4) Run the app

python app.py

Then open:

http://127.0.0.1:7860

---

🧭 How to Use the Interface

1. Upload a chest X-ray (or load a sample)

2. Adjust:

   • Max new tokens
   • Layer selection
   • Head selection
   • Or choose mean attention across all layers/heads

3. Click Generate Findings

4. Click any generated word to visualize:

   • Image ↔ text attention
   • Heatmaps
   • Cross-token relationships

---

🧩 Repository Structure

---

🛠️ Troubleshooting

• Blank heatmap → Ensure output_attentions=True in .generate()
• Distorted attention → Check token count = 1024 (32×32)
• Tokenizer errors → Confirm model.decoder.tokenizer is loaded
• OOM on local machine → Reduce max_new_tokens or use CPU-only settings
• Slow inference → CPU mode is intentionally lightweight; GPU recommended for higher throughput

---

🧪 Model Integration Notes

Compatible with any encoder–decoder or vision–language model that:

• Accepts pixel_values

• Returns attentions when calling

  model.generate(..., output_attentions=True)
  

• Provides a decoder tokenizer:

  model.decoder.tokenizer
  

Ideal for research in:

• Medical AI
• Vision–language alignment
• Cross-modal interpretability
• Attention visualization
• Explainable AI (XAI)

---

❤️ Acknowledgments

• Powered by Gradio and Hugging Face Transformers
• Based on and expanded from the Token-Attention-Viewer project 🔗 https://github.com/devMuniz02/Image-Attention-Visualizer
• Created as part of a thesis on efficient and explainable multimodal medical AI