Overview Process Package

The process package provides functionality for pre-processing, DL-model inference, and post-processing in combination with the provided data model. The main building block of this package is the Filter and its associated FilterParams which process Subject instances. Due to the standardized interface, the chaining of multiple filters in a FilterPipeline is feasible, improving clarity and reproducibility. Furthermore, this package provides an invertibility mechanism for filters that implement invertible process steps. This feature renders feasibility to restore the original physical orientation of the processed Image, which may be crucial when processing medical imaging data. However, subsequent data processing with multiple filters limits the invertibility because the data experiences information loss.

This package provides a basic set of extensible filter implementations. Currently, the process package includes exclusively filters often applied in auto-segmentation development. However, we want to encourage the community to implement and share their filters (e.g., via pull requests to the PyRaDiSe GitHub repository). The recommended workflow for implementing new filters is documented in the documentation of the Filter class.