v2.0.0

This release marks a major overhaul of ClinicaDL, refactoring the entire framework into a Python API.

ClinicaDL is no longer a command-line tool. The previous clinicadl ... CLI and all of its
sub-commands have been removed and replaced by a flexible Python API that lets users build and
customize deep learning pipelines from high-level building blocks. The goal is to make
the code more maintainable, scalable, and user-extensible, while preserving the core values of
ClinicaDL: reproducibility and data leakage prevention.

⚠️ Breaking change: Backward compatibility is broken with all 1.x versions. There is no automatic
migration path. Existing CLI commands and scripts must be rewritten with the new Python API. See the
updated documentation and user guide to get started.

Highlights

• Full rewrite of the core library, now object-oriented and modular.
• CLI removed, replaced by Python API.
• Native BIDS reading support.
• Interoperability with popular deep learning tools for medical imaging: PyTorch, MONAI and TorchIO.
• New MAPS architecture for managing model outputs and metadata.
• Configuration-based design with dedicated pydantic config classes, serializable to/from JSON.
• Extensive and updated documentation (API reference, user, installation and contribution guides).
• Drops Python 3.9; supports Python 3.10–3.14.

Added

Core modules:

• clinicadl.data: for building PyTorch objects able to manipulate neuroimaging data.
• clinicadl.transforms: for transforming 3D neuroimaging data.
• clinicadl.split: for splitting data into training, validation and test sets.
• clinicadl.networks: for building neural networks.
• clinicadl.losses: for creating a criterion to minimize during training.
• clinicadl.optim: for configuring optimization during training.
• clinicadl.models: for defining models, which encompass neural networks, loss functions, optimizers, and the training and evaluation logic.
• clinicadl.train: for training and evaluating a model.
• clinicadl.infer: for customizing the inference stage, such as performing post-processing or combining outputs from multiple neural networks.
• clinicadl.metrics: for evaluating models.
• clinicadl.callbacks: for monitoring and customizing the training and evaluation phases.
• clinicadl.io: for manipulating the file directories produced and read by ClinicaDL.

Configuration classes:

• Use of the serialisable pydantic dataclasses to save the configuration of
  an experiment.

Interoperability with other tools:

• Compatible with TorchIO transforms.
• Integration with MONAI metrics.
• Compatible with native PyTorch neural networks, optimizers, LR schedulers
  and loss functions.

Documentation

• Rewritten Sphinx documentation with a complete API reference and new user, installation, and contribution guides.

Testing

• Unit-tests with a 99% test coverage.
• Specific tests for GPU and multi-GPU setups.
• Functional tests covering the API end to end.

Changed

• Internal architecture redesigned around independent modules.
• All training now goes through Trainer.

Removed

• The entire command-line interface, including all clinicadl ... commands (e.g. clinicadl train, clinicadl random-search, clinicadl preprocessing run).
• Synthetic data generation (will be restored in a future release).
• Quality check for Clinica's pipelines (will probably be integrated to Clinica).
• Random search.
• Interpretation commands (will be restored in a future release).
• Support for Python 3.9.

Breaking Changes

• Backward compatibility is broken with all 1.x versions.