Training-free motion correction for cardiac T1 mapping using DINOv3 features.
TL;DR
This study proposes a novel training-free motion correction framework leveraging frozen DINOv3 foundation model features to achieve robust myocardial alignment without task-specific network training or fine-tuning, providing solution for cardiac T1 mapping.
Abstract
Cardiac T1 mapping is susceptible to respiratory motion, particularly due to the substantial contrast variations and signal inversions across different inversion times (TI). This study proposes a novel training-free motion correction framework leveraging frozen DINOv3 foundation model features to achieve robust myocardial alignment without task-specific network training or fine-tuning. For each image pair, dense DINOv3 features extracted via a frozen ViT-S/16+ encoder are projected into a compact pair-wise PCA feature space. We introduce a variance-aware channel selection and adaptive weighting strategy to prioritize structurally reliable features across the TI-dependent contrast fluctuations. The deformation field is optimized through a hybrid loss function combining weighted DINO-based normalized cross-correlation (NCC), an auxiliary image-domain NCC constraint, and smoothness regularization. Evaluated on the public STONE T1 mapping dataset (32 subjects, 1600 image pairs), the proposed method outperforms existing image-based registration baselines, achieving a Dice similarity coefficient (DSC) of 0.839 ± 0.090 and an HD95 of 1.965 ± 1.705. Furthermore, the method yields the highest T1 fitting quality, with a myocardial R2 of 0.982 ± 0.016, while maintaining topology preservation. Our framework provides solution for cardiac T1 mapping, improving structural alignment under intensity variations without requiring large-scale annotated training data.