HistoPath

HistoPath is a full-stack diagnostic support system capable of detecting metastatic cancer in lymph node sections. It combines a custom CNN inference engine for binary classification with a RAG-powered AI assistant to provide explainable, context-aware clinical insights.

---

System Architecture

The system uses a decoupled architecture to separate ML inference from the application logic.

graph TD
    User((User)) -->|HTTPS| UI["Client Layer<br/>(Next.js 16 / React 19)"]
    
    subgraph "Orchestration Layer"
        UI -->|Auth| Auth[Clerk Auth]
        UI -->|API Routes| API[Next.js API Gateway]
    end
    
    subgraph "Inference Service (Python/FastAPI)"
        API -->|POST /predict| Inference[Model Service]
        Inference -->|Tensor| CNN[Custom CNN Model]
        Inference -->|Gradients| GradCAM[Explainability Engine]
    end
    
    subgraph "Knowledge Engine (RAG)"
        API -->|Query| VectorDB[("Pinecone Vector DB")]
        API -->|Context + Prompt| LLM[Groq Inference Engine]
    end
    
    subgraph "Persistence"
        API -->|Write| DB[("PostgreSQL + Prisma")]
    end

Loading

Key Design Decisions

• Hybrid Compute: Heavy image processing is handled by a Python/FastAPI service, keeping the Next.js UI responsive.
• Explainable AI: Grad-CAM visualizes which tissue regions influenced the prediction, helping users understand the model's logic.
• Context-Aware Assistant: The AI assistant uses RAG to ground its answers in real-time analysis data, reducing hallucinations.

Key Features

• Metastasis Detection: High-precision classification of tissue patches from the PCam dataset.
• Interactive Heatmaps: Overlay heatmaps on slides to localize potential tumor regions.
• Clinical Assistant: Chatbot that explains predictions and interprets confidence scores.
• Secure History: Saves analysis sessions and chat logs for future reference.

Tech Stack

Component	Technology	Description
Frontend	Next.js 16 (App Router)	Server-side rendering, React 19, Tailwind CSS.
Inference API	Python / FastAPI	Async handling of model prediction requests.
ML Core	TensorFlow / Keras	Custom CNN architecture trained on PCam.
Image Processing	OpenCV / Pillow	Preprocessing and heatmap superposition.
Database	PostgreSQL + Prisma	Relational data storage with type safety.
Auth	Clerk	Secure user authentication and session management.
LLM / RAG	Groq SDK + Pinecone	Low-latency inference and semantic search.

Dataset & Model

The model detects metastatic vs. non-metastatic tissue using the PatchCamelyon (PCam) dataset (>280k scans).

Disclaimer: For research use only. Not for professional medical diagnosis.

Contributing

Contributions to the CNN architecture or UI are welcome. Please open an issue or PR.

License

MIT License. See LICENSE for more information.