Pixel-wise-predictions with DNNs

• Depends on the problem dependet on structure variability, background variabilty, intensity profile

• Ambiguity in labeling

• Missing data in multimodal cases

• Images may come from different domaines

• 
  

• Problem: Small structures are dominated by larger ones

• DSC=1 if sets perfectly overlap, 0 when no overlap

• Problem: Details get lost because of too many pooling operations

• How to use contextual information and retain high resolution?

• Solution 1: Combination from different scales to retain resolution

• Solution 2: Unet

  • Encoding part (reduction in size)

  • Decoding part (upscaling)

  • Skip connection to retain high resolution information

• MSE or derived loss function —> not capture what makes images perceptually similar

• Advanced loss function: Measure distance between deep features going beyond simple pixel wise difference

—> MSE

• Discriminator identifies whether the input image is real or not

• This image is generated by a generatot network

• Arhcitecture fully connected or Unet structures

• Syntheszising target image from source imahe

• Challenges: Labels and features may not be paired

• Archtitecture for npaired prblem: