How bubblyyi/coronary-artery-tracking-via-3d-cnn-classification works
Main Components
• Data Processing & Preparation:
• This is covered by the “data_process_tools” directory. It contains scripts to generate image patch data (for both centerlines and seed/ostia points) and tools to create spacing information.
• Model Definitions:
• The “models” directory holds the definitions for three networks (centerline_net, ostiapoints_net, seedspoints_net). Represent this as a “Models” component.
• Training Pipelines:
• There are three distinct training modules:
• “centerline_train_tools” (contains training runner and log files for centerline network).
• “seedspoints_train_tools” (handles training for seed points, with its own logs).
• “ostiapoints_train_tools” (handles training for ostia/entrance points).
• Checkpoints & Logs:
• The “checkpoint” directory is used to store model checkpoints for classification, ostia, and seeds networks. Also, various training log files exist in their respective subdirectories.
• This as a “Checkpoints & Logs” component that is referenced by the training components.
• Inference & Vessel Tree Construction:
• The “infer_tools_tree” directory contains tools to build the vessel tree using the trained models. It includes modules for building the vessel tree structure, metrics (like Dijkstra), settings, etc.
• This functionality as the “Inference & Tree Construction” module.
• External Dependencies & Execution:
• Use of PyTorch as the deep learning framework and additional packages listed in “requirement.txt”. While not a system module, include an “External Libraries” or “Dependencies” note.
Relationships and Data Flow
• Start at Data Processing:
• Training Phase:
• Each training module (centerline, seedspoints, ostiapoints) uses its dedicated training tools and saves model weights (checkpoints) in the “Checkpoint” component.
• Inference Phase:
• From “Checkpoints & Logs”, the trained models are loaded into the “Inference & Tree Construction” component.
• Iterative Feedback:
Architectural Patterns & Design Principles
• Modular Design:
• The project is built with separation of concerns in mind. Each submodule (data processing, training, inference) is self-contained.
• Data Pipeline Flow:
• A sequential flow from data collection/preparation to model training and then to inference, reflecting the typical machine learning pipeline.
• High Reusability: