BrainSynth

Brain scan synthesizer.

Settings

Basic information about known types of images, idiosyncracies of how images are encoded (e.g., the way partial volume information is stored, labeling schemes etc.)

brainsynth/constants/__init__.py

Pipelines are configured in

brainsynth/config/synthesizer_builder.py

Dataset configurations for known datasets can be generated using DatasetConfig from

brainsynth/config/dataset.py

Building Synthesizer Pipelines

Synthesizer instances are initialized with a SynthBuilder object which defines all the actions performed on images and surfaces that are fed to the Synthesizer. The synthesizer applies a sequence of transformations to each input (e.g., an image or surface) and returns the result. The inputs, however, are only available (known) at runtime. To work around this, we provide a Pipeline class which, when called by a synthesizer, is provided with a dictionary of input images, surfaces, state etc. referred to as mapped_inputs. To select variables from this dictionary, we provide a series of Select* classes (e.g., SelectImage, SelectState). Consequently, these must be used as part of a Pipeline and will typically be the first entry in the list of transformations provided by a particular pipeline (e.g., to select a T1-weighted image).

Sometimes we need to initialize transformations with variables that are only available at runtime. To achieve this, we provde a PipelineModule class which simply defers the initialization of a transformation until runtime.

The hierarchy is something like this

Synthesizer
    PipelineBuilder
        Pipeline
            SubPipeline
                Transforms
            PipelineModule
                SelectImage, SelectState, etc.
                Transforms
            Transforms

A Subpipeline allows you to initialize parts of a pipeline only once although it is used as part of several pipelines.

TODOs (known issues)

• If input contains a batch dimension, it will be removed and re-added. However, even if no batch dimension is present, it will still be added in the end. I should fix that...
• Check LR flip. Are segmentation labels correctly flipped/handled?
• Add possibility of flipping along any axis?
• photo_mode and exvivo are both untested.