Normalized Total Gradient for Contrast-Robust 3D Medical Image Registration

Yimin Luo, Yuheng Fan, Heejong Kim, He Zhao, Mert R. Sabuncu, Qingyu Zhao

Keywords: Affine medical image registration, Contrast, Normalized total gradient (NTG), Vision transformer (ViT).

Abstract: Accurate 3D medical image registration remains challenging under varying image contrast and acquisition conditions. Conventional similarity metrics such as normalized cross-correlation (NCC) is highly sensitive to intensity variations, leading to unreliable alignment across modalities. To reduce this limitation, we revisit the core of registration and introduce an objective function for model optimization based on 3D Normalized Total Gradient (NTG). Moreover, we improve registration stability and scalability by introducing a multi-scale dilated attention (MSDA) module. Extensive experiments demonstrate that 3D NTG consistently improves Dice and HD95 scores compared with NCC, and maintains robustness under contrast perturbations. The approach requires no architectural complexity and generalizes across pre-trained backbones and modalities. Our findings highlight the significance of gradient-domain metrics for reliable 3D medical image registration, offering a more practical, interpretable, and contrast-robust solution for clinical imaging scenarios.The code is publicly available at: https://github.com/huanlemin/MDViT-NTG.

Primary Subject Area: Image Registration

Secondary Subject Area: Unsupervised Learning and Representation Learning

Registration Requirement: Yes

Reproducibility: https://github.com/huanlemin/MDViT-NTG

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 236