0.6.0

Added

• isotropic random scaling option
• volume padding with user-specified constant
• subpixel layer for superresolution
• various loss functions for regression (smooth L1 loss, cosine loss etc.)
• handler for early stopping mechanism
• aggregator with multiple outputs including labels in CSV
• nnUNet, an improved version of UNet3D
• data augmentation with mixup and mixmatch
• documentation contents
• demo for learning rate scheduling
• demo for deep boosted regression
• initial integration of NiftyReg Resampler
• initial integration of CSV reader

Fixed

• issue of loading binary values of NIfTI file
• various fixes in CI tests
• prefix name for aggregators
• various improvements in error messages
• issue of batch indices in the conditional random field
• issue of location selection in the weighted sampler
• model zoo: compatibility upgrade
• model zoo: new decathlon hippocampus dataset

Changed

• feature normalisation types options: instance norm, group norm, batch norm
• convolution with padding option
• various documentation and docstrings
• defaulting to remove length one dimensions when saving a 5D volume

0.5.0

Added

• Version controlled model zoo with git-lfs
• Dice + entropy loss function
• Antialiasing when randomly scaling input images during training
• Support of multiple optimisers and gradients in applications

Fixed

• An issue of rounding image sizes when pixdim is specified
• An issue of incorrect Dice when image patch does not include every class
• Numerous documentation issues

Changed

• Tested with TensorFlow 1.12

0.4.0

Added

• niftynet.layer: new layers
  • Tversky loss function for image segmentation
  • Random affine augmentation layer
  • Random bias field augmentation layer
  • Group normalisation layer
  • Squeeze and excitation blocks
• Documentation
  • Image reader and window sampler demo
  • 2D U-net demo
  • Dense CRF layer demo
  • FAQ list
• Misc.
  • Subject id from filename with regular expression
  • Versioning with python-versioneer
  • Tested with TensorFlow 1.10

Changed

• niftynet.engine: improved core functions
  • IO modules based on tf.data.Dataset (breaking changes)
  • Decoupled the engine and event handlers
• Migrated the code repository, model zoo, and niftynet.io source code to
  github.com/niftk.

0.3.0

Added

• Support for 2D image loading optionally using skimage, pillow, or simpleitk
• Image reader and sampler with tf.data.Dataset
• Class-balanced image window sampler
• Random deformation as data augmentation with SimpleITK
• Segmentation loss with dense labels (multi-channel binary labels)
• Experimental features:
  • learning-based registration
  • image classification
  • model evaluation
  • new engine design with observer pattern

Deprecated

• Deprecating the CmicLab repository in favour of GitHub

0.2.2

Added

• Improvements for running validation iterations during training

Fixed

• Bugs when running validation iterations during training
• Minor bugs in loss function modules, histogram standardisation, user parameter parsing

0.2.1

Added

• Support for custom network / application as external modules
• Unified workspace directory via global configuration functionalities
• Model zoo for network / data sharing
• Automatic training / validation / test sets splitting
• Validation iterations during training
• Regression application
• 2D / 3D resampler layer
• Versioning functionality for better issue tracking
• Academic paper release: "NiftyNet: a deep-learning platform for medical imaging"
• How-to guides and a new theme for the API and examples documentation

0.2.0

Added

• Support for unsupervised learning networks, including GANs and auto-encoders
• An application engine for managing low-level operations required by different types of high-level applications
• NiftyNet is now available on the Python Package Index: pip install niftynet
• NiftyNet website up and running: http://niftynet.io
• API reference published online: http://niftynet.rtfd.io/en/dev/py-modindex.html
• NiftyNet source code mirrored on GitHub: https://github.com/NifTK/NiftyNet
• 5 new network implementations:
  1. DenseVNet
  2. HolisticNet
  3. SimpleGAN
  4. SimulatorGAN
  5. VariationalAutoencoder (VAE)

Fixed

• Bugs (30+ issues resolved)

0.1.1

Added

• Source code open sourced (CMICLab, GitHub)
• Initial PyPI package release
• Refactored sub-packages including engine, application, layer, network
• Command line entry points
• NiftyNet logo

Fixed

• Bugs in data augmentation, I/O, sampler