Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Strategies for Robust Deep Learning Based Deformable Registration

Joel Honkamaa, Pekka Marttinen

TL;DR

The paper tackles the limited generalization of deep learning–based deformable registration across unseen contrasts and modalities by applying a modality-independent input transformation (MIND) before a diffeomorphic registration backbone (SITReg) while training with intra-modality NCC loss. An ensemble of models and targeted training refinements (augmentation, diffusion regularization, NDV, and group-consistency losses) are used to boost robustness without sacrificing diffeomorphic guarantees. The approach preserves in-domain performance while substantially improving multi-modal registration, achieving leading performance on the LUMIR 2025 leaderboard prior to submission. This offers a simple, practical strategy for cross-modality registration that can be extended to other datasets and imaging modalities.

Abstract

Deep learning based deformable registration methods have become popular in recent years. However, their ability to generalize beyond training data distribution can be poor, significantly hindering their usability. LUMIR brain registration challenge for Learn2Reg 2025 aims to advance the field by evaluating the performance of the registration on contrasts and modalities different from those included in the training set. Here we describe our submission to the challenge, which proposes a very simple idea for significantly improving robustness by transforming the images into MIND feature space before feeding them into the model. In addition, a special ensembling strategy is proposed that shows a small but consistent improvement.

Strategies for Robust Deep Learning Based Deformable Registration

TL;DR

The paper tackles the limited generalization of deep learning–based deformable registration across unseen contrasts and modalities by applying a modality-independent input transformation (MIND) before a diffeomorphic registration backbone (SITReg) while training with intra-modality NCC loss. An ensemble of models and targeted training refinements (augmentation, diffusion regularization, NDV, and group-consistency losses) are used to boost robustness without sacrificing diffeomorphic guarantees. The approach preserves in-domain performance while substantially improving multi-modal registration, achieving leading performance on the LUMIR 2025 leaderboard prior to submission. This offers a simple, practical strategy for cross-modality registration that can be extended to other datasets and imaging modalities.

Abstract

Deep learning based deformable registration methods have become popular in recent years. However, their ability to generalize beyond training data distribution can be poor, significantly hindering their usability. LUMIR brain registration challenge for Learn2Reg 2025 aims to advance the field by evaluating the performance of the registration on contrasts and modalities different from those included in the training set. Here we describe our submission to the challenge, which proposes a very simple idea for significantly improving robustness by transforming the images into MIND feature space before feeding them into the model. In addition, a special ensembling strategy is proposed that shows a small but consistent improvement.
Paper Structure (11 sections, 2 equations, 1 figure, 1 table)

This paper contains 11 sections, 2 equations, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. Background
  3. Methods
  4. Backbone
  5. Input transformation
  6. Ensembling
  7. Further details
  8. Results
  9. Discussion
  10. Acknowledgments.
  11. Disclosure of Interests.

Figures (1)

  • Figure 1: Overview of the proposed main idea. The input images go through the MIND transformation before being fed to the registration network. As a result, the network learns to do multi-modal registration even though it is trained with an intra-modality similarity loss. The similarity loss is normalized cross-correlation. Also note that in practice the registration network predicts the deformation in both directions, and the losses are also computed for both directions.