DINOMotion: DINOv2 MRI Motion Tracking

• DINOMotion is a DINOv2-based framework that tracks tissue motion in MRI-guided radiotherapy using feature matching and analytical registration.
• It employs frozen DINOv2 features with LoRA adaptation to robustly detect landmarks under large inter-frame displacements and intensity shifts.
• The approach enhances interpretability with visualized correspondence maps, enabling immediate assessment and reduced reliance on iterative optimization.

Searching arXiv for the DINOMotion paper and closely related references. DINOMotion most specifically denotes a DINOv2-based framework for tissue motion tracking in 2D-Cine MRI-guided radiotherapy, where corresponding landmarks are detected between sequential images and used to derive rigid, affine, or nonlinear registration through closed-form, differentiable solvers rather than iterative per-pair optimization. In the supplied arXiv literature, the name is also used in a broader DINO-based motion context for visual terrain-conditioned dynamics modeling in high-speed autonomous off-road driving; however, the primary, explicitly titled use is the radiotherapy framework reported in "DINOMotion: advanced robust tissue motion tracking with DINOv2 in 2D-Cine MRI-guided radiotherapy" (Salari et al., 14 Aug 2025).

1. Problem setting and conceptual basis

In MRI-guided radiotherapy, real-time 2D-cine imaging allows monitoring of organ motion during dose delivery. The central technical problem is frame-to-frame estimation of organ displacement and deformation under free-breathing or other physiological motion, with sufficient accuracy and sufficiently low latency to support gating and margin reduction. DINOMotion addresses this setting by combining powerful self-supervised feature representations from the DINOv2 Vision Transformer with closed-form analytical registration, thereby avoiding the per-frame-pair iterative optimization characteristic of ANTs with SyN, B-spline registration, NiftyReg, and similar approaches (Salari et al., 14 Aug 2025).

The method is positioned around two explicit limitations of prior approaches. First, existing methods are described as sensitive to large inter-frame displacements, particularly when gradient-based solvers encounter local minima. Second, iterative optimization imposes latency that can preclude real-time tracking. DINOMotion therefore frames robust feature matching and analytical registration as a joint alternative: DINOv2 supplies descriptors that are robust under large rotations or intensity shift, while the registration stage solves for parametric transforms directly at test time.

A key methodological consequence is interpretability. Rather than only regressing a dense deformation field, DINOMotion automatically detects corresponding landmarks to derive optimal image registration, and the correspondences themselves can be visualized. This makes the motion estimate inspectable in a way that is uncommon in purely intensity-based or end-to-end warp-prediction pipelines.

2. Architecture: DINOv2 backbone, LoRA adaptation, and landmark matching

At the core of DINOMotion is a Vision Transformer pretrained with the DINOv2 self-supervised objective. Given an MRI frame $I$, the model produces multi-scale feature maps

$$\bigl\{f^{(\ell)}(I)\in\mathbb{R}^{C_{\ell}\times H_{\ell}\times W_{\ell}}\bigr\}_{\ell=1}^{L},$$

where each spatial location acts as a descriptor. The supplied description characterizes these descriptors as invariant to local contrast changes and robust under large rotations or intensity shift, which is the basis for matching across sequential 2D-cine MRI frames (Salari et al., 14 Aug 2025).

The pretrained DINOv2 weights are frozen, and trainable capacity is introduced through Low-Rank Adaptation. If a linear layer is parameterized by $W\in\mathbb{R}^{d\times k}$, LoRA introduces

$W' = W + \Delta W,\quad \Delta W = A\,B,$

with $A\in\mathbb{R}^{d\times r}$, $B\in\mathbb{R}^{r\times k}$, and $r\ll\min(d,k)$. The number of trainable parameters per layer becomes $r(d+k)$ instead of $dk$. Only the LoRA matrices $\{A,B\}$ and a small bias term are updated during fine-tuning.

Correspondences are constructed from feature similarity. For a chosen scale $$\bigl\{f^{(\ell)}(I)\in\mathbb{R}^{C_{\ell}\times H_{\ell}\times W_{\ell}}\bigr\}_{\ell=1}^{L},$$0, each spatial location $$\bigl\{f^{(\ell)}(I)\in\mathbb{R}^{C_{\ell}\times H_{\ell}\times W_{\ell}}\bigr\}_{\ell=1}^{L},$$1 in the moving image and $$\bigl\{f^{(\ell)}(I)\in\mathbb{R}^{C_{\ell}\times H_{\ell}\times W_{\ell}}\bigr\}_{\ell=1}^{L},$$2 in the fixed image forms an entry in the similarity matrix

$$\bigl\{f^{(\ell)}(I)\in\mathbb{R}^{C_{\ell}\times H_{\ell}\times W_{\ell}}\bigr\}_{\ell=1}^{L},$$3

For each moving location $$\bigl\{f^{(\ell)}(I)\in\mathbb{R}^{C_{\ell}\times H_{\ell}\times W_{\ell}}\bigr\}_{\ell=1}^{L},$$4, the best match is

$$\bigl\{f^{(\ell)}(I)\in\mathbb{R}^{C_{\ell}\times H_{\ell}\times W_{\ell}}\bigr\}_{\ell=1}^{L},$$5

This produces landmark pairs $$\bigl\{f^{(\ell)}(I)\in\mathbb{R}^{C_{\ell}\times H_{\ell}\times W_{\ell}}\bigr\}_{\ell=1}^{L},$$6, after which a RANSAC-style filter may be applied to remove outliers.

3. Registration models and analytical inference

Once correspondences are available, DINOMotion solves for a parametric mapping

$$\bigl\{f^{(\ell)}(I)\in\mathbb{R}^{C_{\ell}\times H_{\ell}\times W_{\ell}}\bigr\}_{\ell=1}^{L},$$7

that aligns moving landmarks to fixed landmarks. The common optimization form is

$$\bigl\{f^{(\ell)}(I)\in\mathbb{R}^{C_{\ell}\times H_{\ell}\times W_{\ell}}\bigr\}_{\ell=1}^{L},$$8

with the crucial property that each supported transformation family admits a one-shot analytical solver (Salari et al., 14 Aug 2025).

For rigid registration, the transform is $$\bigl\{f^{(\ell)}(I)\in\mathbb{R}^{C_{\ell}\times H_{\ell}\times W_{\ell}}\bigr\}_{\ell=1}^{L},$$9, where $W\in\mathbb{R}^{d\times k}$0 and $W\in\mathbb{R}^{d\times k}$1. The translation aligns centroids, and the rotation is obtained from the orthogonal Procrustes problem via SVD of the covariance matrix. For affine registration, the transform is $W\in\mathbb{R}^{d\times k}$2, $W\in\mathbb{R}^{d\times k}$3, with parameters given by least squares over the landmark pairs. For nonlinear registration, DINOMotion uses thin-plate spline: $W\in\mathbb{R}^{d\times k}$4 with a regularization parameter $W\in\mathbb{R}^{d\times k}$5 that trades off bending energy versus affine behavior.

This analytical formulation determines the runtime behavior. At test time, DINOMotion performs one forward pass through frozen DINOv2 plus LoRA adapters, computes the similarity matrix and extracts $W\in\mathbb{R}^{d\times k}$6 correspondences, and then solves for $W\in\mathbb{R}^{d\times k}$7 via SVD or matrix inversion. No iterative gradient-based optimization is performed per pair. The reported average inference time is approximately $W\in\mathbb{R}^{d\times k}$8 ms per frame-pair on an NVIDIA RTX 3080, which is described as meeting the real-time requirement of approximately $W\in\mathbb{R}^{d\times k}$9 Hz for MRI-guided radiotherapy.

4. Training protocol, data, and evaluation methodology

The reported training data comprise 10 healthy volunteers and 23 patient scans, with anatomy including pancreas, liver, kidney, and lung, acquired with 2D cine MRI under free-breathing. Preprocessing included ROI cropping, intensity normalization to $W' = W + \Delta W,\quad \Delta W = A\,B,$0, and uniform down-sampling to $W' = W + \Delta W,\quad \Delta W = A\,B,$1 (Salari et al., 14 Aug 2025).

The optimization objective combines a landmark alignment loss,

$W' = W + \Delta W,\quad \Delta W = A\,B,$2

LoRA weight decay $W' = W + \Delta W,\quad \Delta W = A\,B,$3, and the TPS bending regularizer $W' = W + \Delta W,\quad \Delta W = A\,B,$4. The training schedule is specified as 50 epochs, batch size 4, Adam optimizer with initial learning rate $W' = W + \Delta W,\quad \Delta W = A\,B,$5, decayed by $W' = W + \Delta W,\quad \Delta W = A\,B,$6 every 10 epochs, and weight decay $W' = W + \Delta W,\quad \Delta W = A\,B,$7. Only the LoRA adapter matrices are updated.

Evaluation uses Dice similarity coefficient and Hausdorff distance. The definitions are

$W' = W + \Delta W,\quad \Delta W = A\,B,$8

and

$W' = W + \Delta W,\quad \Delta W = A\,B,$9

These metrics are reported for rigid, affine, and nonlinear variants, allowing separation of the contribution of the correspondence engine from the expressivity of the transformation family.

5. Reported performance, interpretability, and clinical implications

The abstract reports Dice scores of $A\in\mathbb{R}^{d\times r}$0 for the kidney, $A\in\mathbb{R}^{d\times r}$1 for the liver, and $A\in\mathbb{R}^{d\times r}$2 for the lung, with corresponding Hausdorff distances of $A\in\mathbb{R}^{d\times r}$3 mm, $A\in\mathbb{R}^{d\times r}$4 mm, and $A\in\mathbb{R}^{d\times r}$5 mm, respectively. The volunteer-level results are also given in a transformation-specific form: kidney DSC $A\in\mathbb{R}^{d\times r}$6 and HD $A\in\mathbb{R}^{d\times r}$7 mm for rigid/affine/nonlinear; liver DSC $A\in\mathbb{R}^{d\times r}$8 and HD $A\in\mathbb{R}^{d\times r}$9 mm; lung DSC $B\in\mathbb{R}^{r\times k}$0 and HD $B\in\mathbb{R}^{r\times k}$1 mm (Salari et al., 14 Aug 2025).

Patient-level results are described as showing the same trend, with nonrigid DINOMotion consistently outperforming NiftyReg, VoxelMorph, and each ANTs variant, especially under large deformations. The runtime comparison is also explicit: approximately $B\in\mathbb{R}^{r\times k}$2 ms per frame-pair for DINOMotion, compared to seconds for ANTs-SyN.

Interpretability is treated as a first-class property rather than a by-product. Because DINOMotion explicitly matches discrete landmarks, one can visualize the correspondence graph by overlaying colored arrows or linking keypoints between frames. The similarity maps $B\in\mathbb{R}^{r\times k}$3 at each scale show which anatomical structures drive each match, providing radiotherapists with intuitive feedback on tracking reliability. This suggests that DINOMotion is intended not only as a registration engine but also as a clinically inspectable decision-support component.

The stated clinical implication is real-time organ motion tracking for 2D-cine MRI-guided radiotherapy, potentially reducing planning margins and sparing healthy tissue. Remaining challenges are also explicit: extending to volumetric $B\in\mathbb{R}^{r\times k}$4D cine streams and handling out-of-plane motion.

6. Secondary usage in off-road dynamics modeling and relation to adjacent DINO-based motion literature

In the supplied literature, “DINOMotion” is also used as the name of an approach in high-speed autonomous off-road driving that predicts changing dynamics induced by terrain as a function of visual inputs. That formulation leverages a pre-trained visual foundation model, DINOv2, extracts patch-wise features of size $B\in\mathbb{R}^{r\times k}$5 per $B\in\mathbb{R}^{r\times k}$6 pixel patch, reduces them by offline PCA from $B\in\mathbb{R}^{r\times k}$7, and then compresses them further through a lightweight MLP encoder $B\in\mathbb{R}^{r\times k}$8 for each wheel. The resulting features are attached to LiDAR points, accumulated in a robot-centric $B\in\mathbb{R}^{r\times k}$9D voxel grid with $r\ll\min(d,k)$0 m voxels using “closest-observation-wins,” and collapsed to a $r\ll\min(d,k)$1D top-down terrain-feature map $r\ll\min(d,k)$2 queried by an MPC planner and a hybrid dynamics model (Gibson et al., 2024).

That dynamics model combines a bicycle model with a learned LSTM compensation term. The state is $r\ll\min(d,k)$3 plus internal delay states, controls are $r\ll\min(d,k)$4, and map inputs are $r\ll\min(d,k)$5. Training uses approximately $r\ll\min(d,k)$6 million five-second trajectory segments, with $r\ll\min(d,k)$7 thousand held-out test segments, collected over more than $r\ll\min(d,k)$8 km across four terrain types. Reported gains include a mean error reduction of about $r\ll\min(d,k)$9 from approximately $r(d+k)$0 m to approximately $r(d+k)$1 m for direct feature input, and a distance-independent encoder that recovers vision gains out to about $r(d+k)$2 m with mean error about $r(d+k)$3 m versus $r(d+k)$4 m baseline.

This secondary usage is methodologically distinct from the radiotherapy framework. In the off-road formulation, DINOv2 is used to condition terradynamics and planning; in the radiotherapy formulation, DINOv2 provides descriptors for landmark correspondence and analytical registration. The shared motif is the use of pretrained visual representations as motion-relevant priors, but the state spaces, objectives, and inference procedures differ substantially.

A further terminological clarification concerns neighboring DINO-based motion papers. MotionAdapter extracts attention-derived motion fields from 3D full-attention modules in DiT-based text-to-video models and applies DINO-guided motion customization for content-aware motion transfer; its domain is video synthesis rather than registration or terradynamics (Zhang et al., 5 Jan 2026). Conversely, the dynamic SLAM paper "D$r(d+k)$5FlowSLAM: Self-Supervised Dynamic SLAM with Flow Motion Decomposition and DINO Guidance" is explicitly described in the supplied details as not using any DINO features and as not introducing a “DINOMotion” concept; instead, it is summarized as a dual-flow self-supervised dynamic SLAM method based on static flow $r(d+k)$6, dynamic flow $r(d+k)$7, and a dynamic update module (Yu et al., 2022). These distinctions are important because the name “DINOMotion” can otherwise be conflated with any DINO-conditioned motion method, even when the underlying task is unrelated.

7. Significance and methodological profile

Across its explicit and implicit usages in the supplied arXiv material, DINOMotion designates a class of motion methods in which pretrained DINO-family visual representations are not merely auxiliary features but the principal mechanism for robustness under appearance variation, misalignment, or projection uncertainty. In the radiotherapy setting, that design is expressed through frozen DINOv2 features, LoRA-based task adaptation, discrete landmark matching, and one-shot rigid, affine, or TPS registration; in the off-road setting, it is expressed through DINOv2 terrain descriptors, PCA compression, projection distance-independent encoding, lightweight mapping, and hybrid learned-parametric dynamics models (Salari et al., 14 Aug 2025).

The radiotherapy instantiation is particularly notable for the way it couples robustness, speed, and interpretability. The framework is described as robust against large misalignments, efficient because only low-rank adapters are trained, and interpretable because explicit visual correspondences are available. A plausible implication is that its methodological identity lies less in any single network component than in the deliberate replacement of opaque dense-warp regression and slow iterative optimization with a feature-matching-plus-analytic-solver pipeline.

The broader literature context supplied here suggests a more general pattern: DINO-derived descriptors can support motion reasoning across very different regimes—organ deformation, vehicle-terrain interaction, and video motion transfer—but the operational meaning of “motion” changes with the downstream solver. In DINOMotion for 2D-Cine MRI-guided radiotherapy, motion is ultimately the image registration map inferred from matched landmarks and solved in closed form.