This is an easy-to-use dataset operation project, which can be used in medical image analysis.
It contains four parts, nnUNet and batchgenerators must be installed first, part of codes comes from nnUNet, thanks to Fabian Isensee et al. for their contributions:
- DatasetCfg: data path, rules of data reading and loading, data augmentation parameters. To use it, a custom dataset config must be created for a certain dataset.
- Dataset: build a dataset class to index a data item, it can read a data item as 2D/3D form automatically based on the param "three_dim" in DatasetCfg.
- Dataloader: an data iteration, which take a Dataset instance and a DatasetCfg instance as its param. It can load a batch of 2D/3D data item automatically according to the param "three_dim" in DatasetCfg.
- DataAugmenter: automatically perform data augmentation for 2D/3D data.
How to use(details can be obtained from the code, it is really really simple): Take Synapse Abdomen dataset as an example,
- 1, build a dataset config class in DatasetCfg, define everything you need, such as data path, etc.
- 2, build a custom Dataset class inherited from Dataset, define the rules of reading a data item (def getitem(self, item))
The data is builded as a dict type through all steps:
- data item: {'data': ndarray, 'seg': ndarray, 'properties': dict, everything you want to keep}
- data batch: {'data': ndarray, 'seg': ndarray, 'indices': list, 'properties': dict, everything you want to keep},