v2.0.0rc1

[2.0.0rc1] – 2025-07-24

This release marks a major overhaul of ClinicaDL, refactoring the entire framework into a modular, API-first design.

Previously centered around the command line, ClinicaDL now provides a flexible Python API that allows users to build and customize deep learning pipelines with high-level configuration objects. The goal is to make the code more maintainable, scalable, and user-extensible, while preserving the core values of ClinicaDL: reproducibility, robust support for neuroimaging, and data leakage prevention.

This is the first release candidate for version 2.0.0, with the final release planned for September 2025.

Highlights

• Full rewrite of the core library — now fully object-oriented and modular.
• New modules (see next section)
• New MAPS architecture for managing model outputs and metadata.
• Clear configuration-based design with JSON files and dedicated config classes based on pydantic.
• Modern deep learning tooling: PyTorch, MONAI, TorchIO, HuggingFace, MLflow, and Weights & Biases support.
• Extensive and fully updated documentation.

Added

Core Modules:

• Trainer: high-level training controller managing full model lifecycle.
• ClinicaDLModel: flexible base class to define and extend custom architectures.
• CapsDataset: redesigned dataset class in the new dataset module, tailored for CAPS/MAPS.
• Splitter: new module for managing train/val/test split logic.
• Maps: a structured and reproducible representation of model outputs and metadata.

Configuration Classes:

• OptimizationConfig, DataloaderConfig, LossConfig, TransformConfig, MetricConfig, etc.
• Designed to be composable and readable using TOML files.
• Stored alongside results to ensure experiment traceability.

Integration with Modern Tools:

• Transforms: use of torchio and monai.transforms for preprocessing and data augmentation.
• Metrics: integration with MONAI metrics and support for custom metrics.
• Networks: fully compatible with native PyTorch models.
• Logging: support for MLflow and Weights & Biases (W&B) out of the box.
• HuggingFace: integration point for loading pretrained models and tokenizers.

Documentation:

• Fully rewritten Sphinx documentation with improved structure and usage examples.
• Interactive object documentation and visual MAPS structure navigation.

Changed

• All pipelines removed and replaced by a unified API-driven interface.
• Internal architecture redesigned for independent modules that can be combined or extended.
• CLI options replaced by TOML configuration — reducing duplication and increasing clarity.
• All training now done through Trainer, using configuration objects and custom hooks.

Removed

• All legacy CLI commands (e.g., clinicadl train, clinicadl random-search, etc.).
• Old pipelines (train_from_json, preprocessing run, etc.).
• JSON-based configuration files.
• Hardcoded command-line flags and argparse logic.

Breaking Changes

• Backward compatibility is broken with all 1.x versions.
• You must migrate to the new API and TOML-based configuration system.