FuseMedML Framework

The fuse.eval package is a standalone library designed for evaluating machine learning models across various performance metrics and facilitating model comparisons. It is designed to be agnostic of the ML library used to generate the evaluation data.

Key capabilities include:

• Massive Data Handling: Automatic batching and multiprocessing for datasets that exceed RAM capacity.
• Multi-class Support: Automatic "one vs. rest" logic for metrics that extend from binary to multi-class.
• Statistical Significance: Built-in support for confidence intervals and model comparison tests (Paired Bootstrap, DeLong's test, McNemar's test).
• Pipelining: Ability to combine sequences of metrics and interleave them with data operations, where metrics can depend on previous results.
• Advanced Analysis: Automatic per-fold evaluation and sub-group analysis.

FuseMedML is organized into several functional modules:

• fuse.data: A declarative data processing pipeline designed for complex multi-modality scenarios. It includes advanced caching with automatic stale cache detection, and ready-to-use Dataset and Sampler classes.
• fuse.eval: A standalone library for evaluating ML models (even those not trained with FuseMedML). It provides metrics, statistical significance tests, calibration, thresholding, and model comparison utilities.
• fuse.dl: Contains reusable deep learning components, including model architectures and loss functions.

The fuse.utils package is organized into several specialized subpackages to support different aspects of ML development:

• fuse.utils.config_tools: Tools for managing configuration.
• fuse.utils.cpu_profiling: Utilities for profiling CPU usage.
• fuse.utils.data: Data manipulation and handling utilities.
• fuse.utils.file_io: File input/output operations.
• fuse.utils.multiprocessing: Support for parallel execution.
• fuse.utils.rand: Randomization utilities.
• fuse.utils.remote_execution: Tools for running tasks on remote systems.
• fuse.utils.tests: Testing utilities.

The fuse.dl.losses package provides a modular framework for managing loss functions in deep learning tasks. It is organized into several specialized modules and subpackages:

Subpackages

• fuse.dl.losses.segmentation: Specialized loss functions for segmentation tasks.

Core Modules

• fuse.dl.losses.loss_base: Contains the base classes and fundamental logic for defining custom losses.
• fuse.dl.losses.loss_default: Provides standard, ready-to-use loss implementations.
• fuse.dl.losses.loss_extract: Utilities for extracting specific values from data structures (like nested dictionaries) to compute losses.
• fuse.dl.losses.loss_warm_up: Implements loss scheduling or warm-up logic.
• fuse.dl.losses.loss_wrap_to_dict: Utilities for wrapping loss outputs into dictionary formats, likely to maintain compatibility with the project's hierarchical data pattern.

The fuse.eval.examples package provides various example modules for different evaluation tasks in medical machine learning. You can use these submodules to understand how to implement evaluation logic for specific domains:

• fuse.eval.examples.examples: General evaluation examples.
• fuse.eval.examples.examples_segmentation: Examples specifically tailored for medical image segmentation tasks.
• fuse.eval.examples.examples_seq_gen: Examples for sequence generation tasks.
• fuse.eval.examples.examples_stats: Examples focused on statistical evaluation.

The fuseimg.datasets package provides specialized modules for accessing various medical and general imaging datasets. Each submodule corresponds to a specific dataset, allowing for standardized data loading and preprocessing within the FuseMedML ecosystem.

Supported dataset modules include:

• cmmd: CMMD dataset
• isic: ISIC dataset
• kits21: KiTS21 dataset
• knight: KNIGHT dataset
• mnist: MNIST dataset
• picai: PI-CAI dataset
• stoic21: Stoic21 dataset

The fuse.dl.models package provides high-level model architectures for medical machine learning tasks. It is organized into subpackages for modular model construction:

• fuse.dl.models.backbones: Contains feature extraction architectures (e.g., CNNs, Transformers).
• fuse.dl.models.heads: Contains task-specific output layers (e.g., classification, segmentation heads).

The package also includes specialized model modules for complex architectures like ensembles, multi-head, multi-stream, and Siamese networks.

The fuse.dl.models.heads package provides various model head implementations for medical imaging tasks, including segmentation and classification. The package is organized into specialized submodules based on the dimensionality and task type:

• fuse.dl.models.heads.common: Contains shared utilities or base classes used across different heads.
• fuse.dl.models.heads.head_dense_segmentation: Implements heads designed for dense prediction tasks like segmentation.
• fuse.dl.models.heads.head_global_pooling_classifier: Implements classification heads that utilize global pooling layers.
• fuse.dl.models.heads.heads_1D: Provides 1D-specific model heads.
• fuse.dl.models.heads.heads_3D: Provides 3D-specific model heads for volumetric medical imaging data.

The fuseimg.utils package provides utility functions for medical imaging tasks. It is organized into subpackages and submodules, primarily focusing on image processing and alignment.

Key components include:

• fuseimg.utils.align: A subpackage dedicated to image alignment tasks.
• fuseimg.utils.image_processing: A submodule containing core image processing utilities.