graph LR
Configuration_Manager["Configuration Manager"]
Data_Pipeline["Data Pipeline"]
Network_Architectures_Model_Zoo_["Network Architectures (Model Zoo)"]
Application_Logic_Loss_Functions["Application Logic & Loss Functions"]
Training_Inference_Engine["Training/Inference Engine"]
Evaluation_Module["Evaluation Module"]
Configuration_Manager -- "configures" --> Data_Pipeline
Configuration_Manager -- "configures" --> Application_Logic_Loss_Functions
Data_Pipeline -- "provides batches to" --> Application_Logic_Loss_Functions
Data_Pipeline -- "provides batches to" --> Training_Inference_Engine
Application_Logic_Loss_Functions -- "defines task with" --> Network_Architectures_Model_Zoo_
Application_Logic_Loss_Functions -- "provides loss to" --> Training_Inference_Engine
Network_Architectures_Model_Zoo_ -- "outputs predictions to" --> Application_Logic_Loss_Functions
Network_Architectures_Model_Zoo_ -- "outputs predictions to" --> Evaluation_Module
Training_Inference_Engine -- "interacts with" --> Network_Architectures_Model_Zoo_
Training_Inference_Engine -- "requests data from" --> Data_Pipeline
Training_Inference_Engine -- "sends outputs to" --> Evaluation_Module
click Configuration_Manager href "https://github.com/CodeBoarding/GeneratedOnBoardings/blob/main/NiftyNet/Configuration_Manager.md" "Details"
click Data_Pipeline href "https://github.com/CodeBoarding/GeneratedOnBoardings/blob/main/NiftyNet/Data_Pipeline.md" "Details"
click Network_Architectures_Model_Zoo_ href "https://github.com/CodeBoarding/GeneratedOnBoardings/blob/main/NiftyNet/Network_Architectures_Model_Zoo_.md" "Details"
click Application_Logic_Loss_Functions href "https://github.com/CodeBoarding/GeneratedOnBoardings/blob/main/NiftyNet/Application_Logic_Loss_Functions.md" "Details"
click Training_Inference_Engine href "https://github.com/CodeBoarding/GeneratedOnBoardings/blob/main/NiftyNet/Training_Inference_Engine.md" "Details"
The NiftyNet deep learning system is structured around a core set of interconnected components designed for medical image analysis. The Configuration Manager acts as the central hub, initializing and distributing system parameters to both the Data Pipeline and the Application Logic & Loss Functions. The Data Pipeline is responsible for preparing raw medical image data into network-ready batches, which it then provides to both the Application Logic & Loss Functions and the Training/Inference Engine. The Application Logic & Loss Functions defines the specific task, leveraging Network Architectures (Model Zoo) for model structures and providing task-specific loss functions to the Training/Inference Engine. The Training/Inference Engine orchestrates the entire deep learning workflow, interacting with Network Architectures (Model Zoo) for model execution, requesting data from the Data Pipeline, and sending processed outputs to the Evaluation Module. Finally, the Evaluation Module quantifies and reports model performance, providing critical insights into the system's effectiveness.
Configuration Manager [Expand]
Centralized control for application settings, user parameters, and global configurations. It initializes the system's operational parameters, including data sources, model specifications, and training/inference options.
Related Classes/Methods:
Data Pipeline [Expand]
Manages the entire data flow from raw input to network-ready batches. This includes loading, partitioning, preprocessing (normalization, augmentation, padding), and sampling (patch/window extraction and batching) of medical image data.
Related Classes/Methods:
niftynet/io/image_sets_partitioner.pyniftynet/io/image_reader.pyniftynet/layer/affine_augmentation.pyniftynet/engine/image_window_dataset.py
Network Architectures (Model Zoo) [Expand]
Houses a collection of pre-defined deep learning models (e.g., UNet, HighRes3DNet, DenseVNet) and fundamental building blocks (convolution, deconvolution layers). It provides reusable and pluggable network structures for various medical imaging tasks.
Related Classes/Methods:
niftynet/network/unet.pyniftynet/network/highres3dnet.pyniftynet/network/dense_vnet.pyniftynet/layer/convolution.py
Application Logic & Loss Functions [Expand]
Defines the high-level logic for specific medical imaging tasks (e.g., segmentation, classification, GANs). It connects the data pipeline, network architectures, and incorporates appropriate loss functions to guide model training.
Related Classes/Methods:
niftynet/application/segmentation_application.pyniftynet/application/classification_application.pyniftynet/application/base_application.pyniftynet/layer/loss_segmentation.py
Training/Inference Engine [Expand]
The core orchestrator responsible for executing the deep learning workflow. It manages TensorFlow sessions, handles model checkpoints, computes gradients during training, and drives both training and inference loops.
Related Classes/Methods:
niftynet/engine/application_driver.pyniftynet/engine/application_iteration.pyniftynet/engine/handler_gradient.pyniftynet/engine/handler_model.py
Quantifies and reports the performance of trained models using various metrics. It processes model outputs and ground truth to provide insights into accuracy and effectiveness.
Related Classes/Methods: