TumorX

Inspiration

Brain tumors tragically claim thousands of lives each year, and in countries like India, the shortage of radiologists (about 1 per 100,000 people) makes timely diagnosis extremely difficult. Long waiting times for MRI reports, especially in rural and semi-urban areas, motivated us to create TumorX—a system that combines AI, privacy-preserving federated learning, and explainable diagnosis to bridge the gap between medical need and available expertise.

What it does

TumorX is an AI-powered assistant for early tumor detection from MRI scans.

For Patients → Provides a simple report: Tumor Present/Not Present, urgency score (Low / Medium / High), and next steps.
For Doctors → Generates explainable Grad-CAM heatmaps, tumor metrics (volume, size, centroid), growth projection, and exportable reports for treatment planning.

How we built it

MRI Input: Accepts scans .
Preprocessing & QIP: Images resized, normalized, and enhanced with Quantum Image Processing (QIP) to highlight subtle anomalies.
AI Model: A hybrid deep learning model with VGG16 backbone classifies into four categories (Glioma, Meningioma, Pituitary, Healthy) and segments affected regions.
Federated Learning: Training occurs locally at hospitals; only encrypted weights are shared, aggregated into a global model.
- Communication cost:
  [ O(n \cdot d) ]
  where (n) = number of clients and (d) = dimensionality of model updates.
Explainability: Grad-CAM generates visual heatmaps for trust and interpretability.
Dashboards:
- Patient Dashboard → Easy-to-understand results.
- Doctor Dashboard → Detailed analytics with reports.

Our stack: Spring Boot + FastAPI (backend), MongoDB (database), and Flower (federated learning).

Challenges we ran into

Federated Learning Complexity → Handling non-IID (non-identically distributed) hospital data, heterogeneous devices, and high communication overhead.
Security & Privacy → Even with federated learning, model updates risk leaking data; we integrated homomorphic encryption for stronger protection.
Explainability → Building trust was hard since most AI models are black boxes; Grad-CAM was essential for transparency.
Performance Balance → Ensuring accuracy without sacrificing speed or accessibility across different hospital setups.

Accomplishments that we're proud of

Built a hybrid deep learning + federated learning pipeline for secure tumor detection.
Integrated explainable AI with visual heatmaps, enabling trust among doctors.
Ensured patient privacy while still improving a global model collaboratively.
Created dual dashboards tailored for both patients and doctors.

What we learned

Medical AI adoption depends not just on accuracy but on trust, privacy, and accessibility.
Federated learning can solve privacy issues but introduces system heterogeneity challenges.
Explainability is as important as performance in healthcare AI.
The parallels between model confidence and spontaneity in thermodynamics, where:

[ \Delta G^\circ = -RT \ln K_{eq} ]

show that just like negative (\Delta G^\circ) indicates spontaneous reactions, higher model confidence indicates more reliable predictions.

What's next for TumorX

Pilot Deployment in partner hospitals for real-world validation.
Digital Twin Integration → simulate tumor growth and treatment responses.
Blockchain + QKD-secured Federated Learning → for tamper-proof, quantum-safe data exchange.
Neuromorphic Chip Deployment → faster, energy-efficient AI inference at hospitals.
Expand beyond brain tumors to other oncological and neurological disorders.