Model overview

This is an overview of the currently supported restoration models that are tailored to commonly-used imaging scenarios:

csbdeep.models.CARE

Description:
  • Standard model that learns a mapping from source (degraded) to target (restored) images.
  • Source/target can be (multi-channel) 2D or 3D stacks.
  • Expects spatially registered source/target pairs.
Typical use-case:
  • Denoising of live-cell images (e.g. acquired with reduced laser power/exposure).
  • Improving SNR of fast time-lapses (e.g. of vesicle trafficking).
Examples:

csbdeep.models.UpsamplingCARE

Description:
  • Extension of the standard model that will additionally increase sampling along a given (e.g. axial) dimension by a given factor s.
  • Source/target pairs should be registered 3D stacks with the desired pixel size.
  • After training, the model is applied to lower-resolution data producing target stacks with an s-fold increased number of sample planes.
Typical use-case:
  • Improving the axial resolution of volumetric time-lapses when only a limited number of focal planes can be acquired.
Examples:

csbdeep.models.IsotropicCARE

Description:
  • Model that improves axial resolution of (axially) anisotropic stacks.
  • Takes anisotropic 3D stacks as source (important: doesn’t need corresponding target stacks).
  • The PSF of the microscope has to be (approximately) known.
  • Assumes isotropic distribution of biological structures (i.e. don’t use it with highly anisotropic tissue like cortical tissue).
Typical use-case:
  • Enhancing axial resolution of (already acquired) light-sheet microscopy time-lapses of developing embryos.
Examples:

csbdeep.models.ProjectionCARE

Description:
  • Extension of the standard model that will additionally project along a given (default: axial) dimension.
  • While source images should be 3D stacks, the corresponding target images are missing an axis and thus are 2D images.
Typical use-case:
  • Projection and denoising of a 2D manifold (e.g. epithelial tissue) inside a 3D stack.
Examples: