Rigid / Affine Registration of 2D Whole Slide Images (WSIs)

Objective:

Given a pair of images, one Source image and one Target image, in which the Source image is a rigidly transformed version of the Target image. Train a Neural Network to predict the transformation parameters between the two images, or, Find the transformation that maps the Source to the Target image.

$$images = [S, T]$$ $$parameters = [\Delta\Theta, \Delta X, \Delta Y]$$

Dataset:

Pefect Pairs Dataset

• Consists of 8 sets of images.
  • Each image set contains 100 pairs of source and moving images.
  • The source and moving image are the same image, the difference is in the rigid transformation between the images
  • The transformations are known
  • Each set has one info.json file, containing the image pair and the transformation label

Models:

DeepHistReg Inspired Model

• Feature Extractor and Regression Head
  • Feature Extractor consists 6 Forward Blocks
    • Feature Extractor Forward Block: Conv2D() -> BatchNorm() -> PReLU() x2
    • Final Layer of the FE: Conv2D() -> BatchNorm() -> PReLU() -> Conv2D() -> BatchNorm() -> PReLU() -> AveragePool2D()
  • Regression Head
    • Flatten() and Dense(3)
• Input is set of moving and static images
• More information on DHR: dhr.md

Stratified Input Model

• Two inputs, one for each image in the pair
  • 6 Feature Extractor Blocks
  • Final Layer applied to each image seperately
  • Concatentate Source and Moving pipelines
  • Dense(3) output