How cardio-ai/fedkd-for-cardiac-ct works
Key Components to Include:
Data Flow and Processing Layer
• Input Layer
• Local CT scan data storage
• Partial label annotations per hospital
Federated Learning Architecture
• Local Training Components
• CNN Models (Task-specific)
• Knowledge Distillation Pipeline
• Local Data Preprocessing (reference src/data/transforms.py)
• Global Aggregation Layer
• Model Parameter Aggregation
• Federation Coordinator
• Global Model Updates
Model Architecture
• Primary Components
• SWIN-UNETR Transformer (src/models/multihead_swinunetr.py)
• UNet Implementation (src/models/nnunet.py)
• Multi-head Architecture for Different Tasks
• Hinge Points Detection
• Coronary Ostias Detection
• Membranous Septum Detection
• Calcification Detection
Training Pipeline
• Knowledge Distillation Flow
• Task-specific CNN Training
• Transformer Learning from CNN Predictions
• Semi-supervised Learning Component
• Model Conditioning Options (segmentation-based)
Important Elements to Highlight:
Two-Step Semi-Supervised Strategy
• Clear separation between CNN prediction phase
• Knowledge distillation to transformer phase
Multi-Task Learning
• How different tasks are handled by separate heads
• Connection to partial labels from different hospitals
Privacy Preservation
• Model parameter sharing instead of raw data
• Local computation boundaries
Evaluation Flow
• Testing paths for both global and local updates
• Validation pipeline
• Brief descriptions of key components
• Important file references
Additional Notes:
• Keep the federation topology clear and central
• Knowledge distillation pipeline
• Clear boundaries between local and global operations
• Reference to key configuration files (env.yml)
• Integration points for physiological constraints (src/physiological_constraints)