Carotid Plaque Vulnerability Assessment Tool (CDSS)

This project implements a web-based prototype Clinical Decision Support System (CDSS) for the identification and risk stratification of carotid atherosclerotic plaques. Developed in the context of a multimodal research framework, it combines B-mode ultrasound imaging (systolic and diastolic phases) with detailed clinical, laboratory, and biomarker data to assess the vulnerability of carotid plaques. Inference is performed in-browser using a TensorFlow.js model, allowing for real-time prediction without transmitting any patient data to a server.

---

🚀 Features

• Upload systolic and diastolic ultrasound images.
• Input structured clinical data:
  • Demographics (age, gender, smoking status)
  • Medical history (diabetes, hypertension, dyslipidemia, coronary artery disease)
  • Lab results (LDL, HDL, glucose, cholesterol, etc.)
  • Biomarkers (CRP, MMPs, ILs, TNFα, etc.)
• Image and clinical data are preprocessed and passed through a multimodal deep learning model.
• Predicts risk score [0–1] for plaque vulnerability.
• Displays a color-coded risk level: Low, Moderate, or High.
• Highlights key contributing factors and provides personalized clinical recommendations.
• Designed for integration with:
  • ESVS 2023 Clinical Practice Guidelines on Atherosclerotic Carotid and Vertebral Artery Disease (see Supplementary Table \ref{tab:esc_recommendations})
  • AHA Atherosclerotic Plaque Classification (see Supplementary Table \ref{tab:aha_plaque})

---

🧠 Model Architecture

• Image Processing: EfficientNet or similar CNN for systolic and diastolic image inputs.
• Clinical Data Processing: Dense layers for tabular inputs.
• Fusion: Concatenation of image and tabular features followed by dense layers.
• Output: Scalar value ∈ [0, 1] indicating risk of plaque rupture.

---

📁 File Overview

• index.html: User interface for uploading data and displaying results.
• app.js: Handles form logic, image loading, and UI interactivity.
• tf-model.js: Loads the model and performs prediction.
• model/: Place your converted TensorFlow.js model files (model.json and binary weight shards).

---

🛠️ Getting Started

1. Convert your trained model to TensorFlow.js format:

   tensorflowjs_converter --input_format=tf_saved_model path/to/saved_model path/to/model

2. Place the converted model in the /model/ directory.

3. Open index.html in a modern web browser (no backend/server required).

4. Upload ultrasound images and clinical data to receive a vulnerability risk assessment.

📌 Notes

This tool is in the Alpha phase and is for research use only.

It is not approved for clinical diagnosis or therapeutic decision-making.

Predictions must be interpreted in the context of full clinical evaluation.

No patient data is stored or transmitted; all inference occurs client-side.

👨‍⚕️ Intended Users

• Vascular surgeons

• Neurologists

• Interventional radiologists

• Cardiovascular researchers

• Graduate students working on atherosclerosis, medical imaging, or clinical AI

📄 License

This code is intended for academic, non-commercial research use only. Please contact the author for collaboration or licensing inquiries.

✍️ Citation

If you use this tool or its concepts in your research or publication, please cite the following:

BibTeX:

@mastersthesis{stamos2025cdss,
  author = {Evangelos Stamos},
  title = {Deep multimodal fusion of image and non-image data in identification of high-risk carotid atheromatous plaque},
  school = {National Technical University of Athens},
  year = {2025}
}