Adaptive Bidirectional Displacement for Semi-Supervised Medical Image Segmentation

Table of Contents

Overview

Paper: Chi et al., Adaptive Bidirectional Displacement for Semi-Supervised Medical Image Segmentation (cvpr2024 open access or arxiv).

img

(Figures and tables in this post are from the original paper)

Novelties of the Paper

  • They proposed an “Adaptive Bidirectional Displacement” (ABD) method to solve the propbems of consistency learning on semi-supervised medical image segmentation.
  • If the consistency learning utilizes multiple perturbations, it’s learning process will easily get uncontrollable.
  • The ABD was introduced to make it enable to employ mutiple perturbations.
  • ABD has two perturbatons. Image argumentation proceses are used as the image purturbation, resulting weak and strong augmentations. In addition, two networks are also utilized as the network perturbation.
  • ABD framework consists of two types of displacement methods, ABD-R and ABD-I.
  • ABD-R is an adaptive bidirectional displacement with reliable confidence. The lowest confidence patches in the augmented image are displaces with the most reliable regions in the other augmented image.
  • ABD-I is an ABD with inverse confidence. The highest confidence regions in the augmented image are replaced by the lowest reliability patches in the other augmented image.
  • Displaced images are treated as new samples.
  • ABC approach an be added to exsisting methods.

Translated with DeepL (https://www.deepl.com/app/?utm_source=ios&utm_medium=app&utm_campaign=share-translation)

Performance Evaluation Methods

  • Two dataset are used: ACDC and PROMISE12.
  • They evaluated how ACD can improve the existing models, Cross Teaching and BCP, and compare them to other models including U-Net, DTC, URPC, MC-Net, SS-Net, SCP-Net, Cross Teaching and BCP.

img

img

Discussions

  • ABD can improve performance of existing methods and achieved the best results.
  • They examined that it’s important to obtain the best performance of ABD to use all three compornents: image purturbation, ABD-R and ABD-I.