Loading...

Medical Image Registration using Self-Supervised Learning

Kalbasi, Mohammad | 2025

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57983 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Fatemizadeh, Emadoddin
  7. Abstract:
  8. Image registration is fundamental in medical imaging, enabling the accurate alignment and analysis of images from different modalities. This process is essential for various applications, including the assessment of tissue growth and tumor evolution, preoperative and intraoperative planning, and segmentation. While recent advances in deep learning have improved unsupervised monomodal medical image registration, these methods typically rely on transforming images and computing a similarity loss function between the transformed and target images. However, for multimodal image registration, traditional similarity functions often lack expressiveness and are prone to numerous local optima. To address these limitations, recent approaches have leveraged deep learning networks to learn intermediate representations for calculating similarity based on embeddings. While these methods can be effective, there is no guarantee that the learned representations will generalize well to challenging modalities, such as PET and MRI, particularly when the network has not been explicitly trained on these modalities. In this work, we propose a novel two-stage framework that integrates representation learning with unsupervised image registration to enhance generalization and performance across diverse imaging modalities. The first stage employs traditional similarity loss functions for unsupervised multimodal image registration. The second stage introduces a self-supervised training loop, enabling the network to extract embeddings that are invariant to deformations and intensity variations. This dual training strategy ensures robust registration across various imaging conditions while promoting generalization. Our approach was evaluated on the challenging task of registering MRI to PET images. The results demonstrate that the proposed method outperforms conventional multimodal image registration techniques, achieving DICE scores of 72.36% for FLAIR to PET registration, 71.93% for T1w to PET registration, and 82.37% for T1w to FLAIR registration
  9. Keywords:
  10. Image Registration ; Medical Images ; Deep Learning ; Representation Learning ; Self-Supervised Learning ; Medical Image Registration

 Digital Object List

 Bookmark

No TOC