• 

  INTERFACE

• 

  INTERFACE

• 

  INTERFACE

• 

  PATIENT HISTORY

• 

  UPLOAD IMAGE

• 

  DASHBOAD

Inspiration

Pulmoscan IA was inspired by a critical healthcare challenge: the late detection of lung cancer, especially in low-resource environments. Limited access to medical imaging tools, specialists, and stable internet connectivity significantly delays early diagnosis.

The goal was to design an intelligent solution that works anywhere, even offline, while leveraging cloud capabilities when available.

What it does

Pulmoscan IA is an AI-powered assistant that: Analyzes lung medical images (X-rays / CT scans) Detects potential abnormalities Generates a probability score Provides visual explanations (Grad-CAM)

It operates in:

🌐 Online mode (advanced analysis via API) 📱 Offline mode (lightweight local inference)

⚠️ The system does not provide medical diagnosis. It is strictly a decision-support tool.

How we built it

The system is based on a hybrid architecture:

Backend

API built with FastAPI Handles image processing and inference

AI Models

CNNs (DenseNet / MobileNet) for detection Lightweight optimized model for offline Advanced models for online mode

Processing Pipeline

Image acquisition Preprocessing (resize, normalization) Model inference Output generation (score + Grad-CAM heatmap)

Deployment cloud run

Challenges we ran into

Hardware constraints (≥ 8GB RAM) No real dataset (simulation-based approach) Balancing offline performance and accuracy Ensuring reliability without false interpretations Making results interpretable with Grad-CAM

Accomplishments that we're proud of

Functional online + offline AI system Integration of Grad-CAM for explainability Optimization for low-resource environments Scalable and robust architecture Strong ethical safeguards (no diagnosis)

What we learned

Building resilient AI systems for real-world constraints Importance of Explainable AI in healthcare Trade-offs between performance and accuracy Designing hybrid (edge + cloud) architectures Ethical responsibility in medical AI

What's next for Pulmscan IA

Multimodal models (image + text) Better offline optimization (quantization, ONNX) Mobile app development (Android) Patient monitoring features Validation with real medical data Improved UX and workflow automation

Built With

• cloud-firestore
• et
• fastapi
• firebase-suite-(auth
• flutter-(dart)
• gemini-api
• grad-cam
• ollama-(inference-deconnectee-llama-3-vision)
• python
• storage)