Project Overview
The Federated Leukemia Diagnostics System represents a paradigm shift in clinical AI by combining:
- Multimodal AI (Late-Fusion): Dual-input networks processing visual and tabular data simultaneously.
- Federated Learning: Institutions train independently, sharing only model weights to create a master global model.
- Privacy-by-Design: Zero patient data leaves institutional premises, secured further by Differential Privacy.
The Problem We Solve
Traditional ML requires moving raw, sensitive patient data to central servers, risking HIPAA violations and data breaches.
Our solution brings the model to the data. Institutional autonomy is preserved while achieving a global accuracy of 99.2%.
End-to-End Data Flow & Prediction Pipeline
Data Ingestion
User uploads Smear Image & inputs 9 Blood Report values
Preprocessing
Image resized to 224x224. Tabular data standardized via Scaler.
Model Inference
Data passes through CNN + MLP branches & fuses for Sigmoid output
XAI & Results
Diagnosis rendered alongside Grad-CAM heatmaps & Feature impacts
Model Architecture Blocks
Image Input
(224, 224, 3)
Conv2D (8) + L2
MaxPool + Dropout(0.3)
Conv2D (16) + L2
MaxPool + Dropout(0.35)
Global Average Pooling
Image Features
Dense(32) Vector
Clinical Tabular Input
(9 features)
Dense (16) + L2
ReLU + Dropout(0.4)
Clinical Features
Dense(8) Vector
Late Fusion (Concatenate)
(32 + 8 = 40 dimensional vector)
Dense (16) + ReLU + Dropout(0.4)
Dense (1) + Sigmoid
Leukemia Probability Output [0,1]