Diagnoa

Resilient AI for Reliable Diagnosis – Defending Medical Imaging Against Adversarial Attacks

Comment

🚀 Diagnoa – Adversarially Robust AI for Inclusive Healthcare

UN SDG 10: Reduced Inequalities

🌟 Inspiration

Access to timely and accurate healthcare remains a luxury for many. Especially in rural and low-income regions, diagnostic errors and delays lead to life-threatening consequences. While AI has the potential to bridge this gap, it often fails when exposed to real-world distortions like low-quality images or noise.

We were inspired by this digital health divide and motivated to build a system that could offer equal healthcare opportunities—regardless of the environment or input quality.

🩺 What It Does

Diagnoa is a dual-model AI diagnostic system designed to detect pneumonia from chest X-ray images and remain reliable under adversarial attacks. Here's what it delivers:

  • Accepts user-uploaded X-ray images
  • Predicts disease using:
    • A standard AI model
    • An adversarially-trained robust model
  • Visualizes how adversarial noise affects predictions
  • Compares both models’ outputs with confidence scores
  • Ensures fairness by resisting adversarial manipulations that could otherwise lead to misdiagnosis

🛠️ How We Built It

We built Diagnoa using:

  • TensorFlow/Keras: To train two convolutional neural networks—one standard, one robust
  • FGSM (Fast Gradient Sign Method): To generate adversarial examples for robust training
  • Streamlit: For an intuitive, browser-based interface
  • Matplotlib & Seaborn: For data visualizations and confusion matrix plots
  • Google Drive Integration: For dynamic model downloads during app runtime

The architecture includes:

Advanced Systems Architecture and Security Mechanisms

⚠️ Challenges We Ran Into

  • Balancing Accuracy and Robustness: Adversarial training often compromises clean image accuracy. Tuning this trade-off was key.
  • Model Size & Deployment: Compressing models for seamless Streamlit integration without losing performance.
  • Generating Meaningful Adversarial Samples: Ensuring the added noise simulates real-world image degradation.
  • UX Design: Designing a layout that clearly communicates both model results to a non-technical audience.

🏆 Accomplishments We're Proud Of

  • Successfully created two models: one that performs well on clean images and one that maintains accuracy even on adversarial images.
  • Built an interactive platform where users can visually compare how AI robustness affects prediction reliability.
  • Aligned our work directly with SDG 10 by focusing on reducing healthcare inequality through inclusive technology.

📚 What We Learned

  • AI systems must be evaluated not just by accuracy, but also by fairness, robustness, and resilience.
  • Adversarial machine learning isn't just academic—it has powerful implications for real-world safety and access.
  • A good solution is not just technical—it’s also user-friendly, scalable, and empathetic to diverse user needs.

🔮 What’s Next for Diagnoa

  • Disease Expansion: Train models on other conditions like COVID-19, tuberculosis, or lung cancer.
  • Mobile & Offline Compatibility: Optimize for low-resource deployment in remote clinics and villages.
  • Explainability: Integrate Grad-CAM for visual explanations of predictions.
  • Local Partnerships: Collaborate with public health organizations to pilot Diagnoa in underserved areas.
  • Open Source Release: Share datasets, models, and the Streamlit app to invite global contributions.

Built With

Share this project:

Updates