---
title: 'LGANet++: Unsupervised 3D Image Registration'
url: https://www.emergentmind.com/topics/lganet
type: topic
---

# LGANet++: Unsupervised 3D Image Registration

Searching arXiv for the specified paper and closely related registration work.
arxiv_search.query({"search_query":"id:2601.14337 OR ti:\"Unsupervised Deformable Image Registration with Local-Global Attention and Image Decomposition\"","start":0,"max_results":5})
I found the target paper on arXiv and will use it as the primary source. I’ll also retrieve a small set of related registration papers for contextual comparison.
arxiv_search.query({"search_query":"ti:VoxelMorph deformable registration OR ti:PIViT deformable image registration","start":0,"max_results":10})
LGANet++ is an unsupervised deformable image registration framework for medical image analysis that combines local-global attention with image decomposition in a coarse-to-fine architecture to improve registration accuracy, robustness, and generalizability across cross-patient, cross-time, and cross-modal CT–MR scenarios [2601.14337]. It is formulated as a 3D CNN-based registration network with a dual-stream encoder, multi-scale fusion, a local-global attention mechanism, feature interaction and fusion with decomposition, and a diffeomorphic layer intended to produce smooth, topology-preserving deformation fields. In the reported evaluation on five publicly available datasets, the method consistently outperformed several state-of-the-art registration methods, with gains of 1.39% in cross-patient registration, 0.71% in cross-time registration, and 6.12% in cross-modal CT–MR registration; the source code is available at the authors’ public repository [2601.14337].

## 1. Role in deformable image registration

Deformable image registration is described as a critical technology in medical image analysis with applications including disease diagnosis, multi-modal fusion, and surgical navigation. The motivating problem is that traditional methods often rely on iterative optimization, which is computationally intensive and lacks generalizability, while recent deep learning methods with attention-based mechanisms still face difficulty in accurately registering regions with high anatomical variability [2601.14337].

Within that setting, LGANet++ is positioned as an unsupervised, coarse-to-fine deformable registration network. Its defining premise is that registration quality can be improved by jointly modeling long-range contextual dependencies, local correspondences, and structural consistency under intensity variation. This is operationalized through the combination of local-global attention, multi-scale fusion, image decomposition, and diffeomorphic refinement. A plausible implication is that the method is intended not merely as a replacement for a single attention block, but as an integrated design for handling both geometric complexity and modality-dependent appearance change.

## 2. Architectural organization

LGANet++ begins with two identical 3D CNN encoders with shared weights that extract four levels of feature pyramids $F_i$ and $M_i$ for $i=1\ldots4$ from the fixed image $I_f$ and moving image $I_m$, respectively [2601.14337]. These pyramid features are fused across scales by the Multi-Scale Fusion Module (MSFM), which projects each feature map to a low dimensionality $c$ with $c=8$, resizes all feature maps to a common spatial size $s_t$ by pooling or interpolation, and multiplies and reconvolves them to produce a fused map $C_i$:
$$
F'_{k,i} =
\begin{cases}
\mathrm{Interp}(\mathrm{Conv}(F_i)), & s_i < s_t,\\
\mathrm{Conv}(F_i), & s_i = s_t,\\
\mathrm{Pool}(\mathrm{Conv}(F_i)), & s_i > s_t,
\end{cases}
$$
$$
C_i = \mathrm{Conv}\bigl(\bigotimes_{k=1}^4 F'_{k,i}\;\otimes\;\bigotimes_{k=1}^4 M'_{k,i}\bigr).
$$

At the coarsest scale, $i=4$, the Local-Global Attention Module (LGAM) takes $F_4$, $M_4$, and $C_4$ and estimates an initial deformation $\phi_4$. At the finer scales $i=3,2,1$, the Feature Interaction and Fusion Module (FIFM) iteratively refines the deformation. The overall hierarchy is explicitly coarse-to-fine:
$$
\phi_i = \mathrm{Up}(2\cdot\phi_{i+1})\circ \phi_i' + \phi_i'.
$$

This progression propagates context from coarse resolutions and then refines finer details at higher resolutions. In structural terms, the network separates global initialization from progressive local correction, which is consistent with the broader design objective of combining robustness to large deformation with sensitivity to fine anatomical detail.

## 3. Attention, fusion, and decomposition mechanisms

Within LGAM, a 3D correlation layer is first applied between $F_4$ and $M_4$, and the resulting representation is concatenated with $C_4$. The combined feature
$$
X=[F_4;C_4;M'_4]
$$
is then processed by a Position Attention Module (PAM):
$$
[Q,K,V] = \mathrm{Conv}_{qkv}(X),\;\;
[q,k,v]=\mathrm{Reshape}(Q,K,V),
$$
$$
U=\mathrm{Conv}_{res}\bigl(\mathrm{Reshape}(\mathrm{Softmax}(qk^T/\sqrt{d_k})\,v)\bigr).
$$
The resulting representation is split into two branches: Global Attention (GA), which applies self-attention across the entire volume, and Local Attention (LA), which applies self-attention within non-overlapping sub-volumes [2601.14337]. Both branches follow the standard transformer attention formulation
$$
Q = P_q(X),\;\; K = P_k(X),\;\; V = P_v(X),
$$
$$
\mathrm{Attention}(Q,K,V) = \mathrm{Softmax}\bigl(QK^T/\sqrt{d_k}\bigr)\,V,
$$
with $X$ reshaped as either $N\times C$ for global attention or $m\times n\times C$ for local attention. Residual connections preserve learned context, and the outputs are fused through an MLP to form $\phi_4$.

The finer-scale refinement is centered on FIFM and the Image Decomposition Module (IDM). After upsampling $\phi_{i+1}$, the model warps the original features to obtain $W_i$. IDM is then applied both to $W_i$ and to $F_i$:
$$
\phi_z, W_z = \mathrm{IDM}(Z).
$$
For $Z=W_i$ and $Z=F_i$, this yields $(\phi_{i1},W_1)$ and $(\phi_{i2},W_2)$. The method then computes cross-warp consistency through
$$
W_{i1}=W_1\circ\phi_{i2},\qquad W_{i2}=W_2\circ\phi_{i1},
$$
and concatenates $C_i$, $I_w$, and $I_f$ weighted by the Channel-Wise Attention Module (CWAM). Finally, the resulting field $\phi_i'$ is summed with STN-warps of $\phi_{i1}$ and $\phi_{i2}$ to produce $\phi_i$.

The stated purpose of IDM is to decouple anatomical structure from intensity variations, thereby improving consistency between the warped and fixed images. This suggests that LGANet++ treats appearance inconsistency not as noise alone but as a factor that should be explicitly separated from structural correspondence, especially in cross-modal settings.

## 4. Diffeomorphic refinement and loss formulation

Once $\phi_4$ is estimated, LGANet++ applies a differentiable exponential layer intended to ensure invertibility and smoothness of the deformation [2601.14337]. The update is given as
$$
\phi^{t-1} = \phi^t + \phi^t\circ \phi^t,\qquad \phi^t = \phi_i/2^t,\; t=7.
$$
This layer is the mechanism by which the network enforces diffeomorphic behavior at the field-construction stage.

Training is unsupervised. The similarity term is based on local normalized cross-correlation (NCC) over a window $\omega$:
$$
\mathrm{NCC}(I_f,I_w) = \sum_{v\in\omega} \frac{(I_f(v)-\mu_f)(I_w(v)-\mu_w)}{\sigma_f\,\sigma_w}.
$$
The optimization objective is
$$
\mathcal{L}_{sim}=1 - \mathrm{NCC}(I_f,\,I_m\circ \phi),
$$
$$
\mathcal{L}_{reg}=\|\nabla \phi\|_2^2,
$$
$$
\mathcal{L} = \mathcal{L}_{sim} + \lambda\,\mathcal{L}_{reg},\; \lambda=1.
$$

A common misconception is that unsupervised registration dispenses with explicit constraints. In LGANet++, unsupervised learning is paired with a specified similarity measure, a deformation smoothness regularizer, hierarchical refinement, and a diffeomorphic layer. The absence of manual annotations does not imply unconstrained field estimation; rather, the supervision is implicit in the image similarity and regularity terms.

## 5. Data, preprocessing, and optimization protocol

The reported experiments use five publicly available datasets spanning three registration scenarios: LPBA for cross-patient brain MRI, IXI for brain MRI atlas-to-patient registration, OASIS as an external test set, Lung CT for exhalation-to-inhalation cross-time registration, and Abdomen CT–MR for cross-modal liver registration [2601.14337]. Images were skull-stripped or clipped, resampled or cropped to volumes of size up to $192\times192\times192$, and normalized to $[0,1]$.

Implementation was in PyTorch on an NVIDIA RTX 4090 GPU with single-precision 3D convolutions. The batch size was 1, except batch 4 for IXI. Optimization used Adam with initial learning rate $1\mathrm{e}{-4}$, $\beta_1=0.99$, $\beta_2=0.999$, linear decay to $1\mathrm{e}{-6}$ after epoch 200, and total training for 300 epochs. The loss settings were $\lambda=1$ and NCC window $\omega=9^3$.

| Scenario | Datasets | Preprocessing / protocol |
|---|---|---|
| Cross-patient | LPBA, IXI, OASIS | Skull-stripped, resampled/cropped, normalized |
| Cross-time | Lung CT | Clipped, resampled/cropped, normalized |
| Cross-modal | Abdomen CT–MR | Clipped, resampled/cropped, normalized |

The protocol indicates that the method was not evaluated on a single homogeneous benchmark. Instead, the experiments were organized to test generalization across within-modality, longitudinal, and cross-modality conditions, including an external test on OASIS after training on IXI.

## 6. Quantitative performance, interpretation, and limitations

LGANet++ was compared against nine state-of-the-art methods, including VoxelMorph, PIViT, GroupMorph, and RDP, using DSC, HD95, TRE, recall, precision, and Negative Jacobian Determinant (NJD) [2601.14337]. The reported results are as follows:

| Task | Main result | Additional metrics |
|---|---|---|
| LPBA cross-patient | DSC 73.52% | HD95 5.10 mm, NJD 0.01% |
| IXI cross-patient | DSC 83.60% | HD95 2.12 mm |
| OASIS external test | DSC 76.70% | HD95 2.71 mm |
| Abdomen CT–MR | DSC 80.28% | HD95 6.43 mm, recall 84.72%, precision 74.62% |
| Lung CT cross-time | DSC 97.61% | TRE 2.02 mm, NJD 0.002% |

The gains over the previous best were reported as 1.39% on LPBA, 0.76% on IXI, 1.39% on the IXI-to-OASIS external test, 6.12% on Abdomen CT–MR, and 0.71% on Lung CT; TRE on Lung CT decreased by 12.9%. All improvements were statistically significant by Wilcoxon $p<0.05$ except where marginal HD95 differences were noted.

The paper attributes the strong performance to four factors: local-global attention in LGAM for voxel-level correspondences and long-range context; multi-scale fusion in MSFM for semantic integration across resolutions; IDM and CWAM in FIFM for structural consistency and adaptive feature weighting; and coarse-to-fine diffeomorphic refinement for robustness with topology preservation. The reported limitations are also specific. The method can exhibit occasional non-diffeomorphic local folds in very challenging cross-modal cases, reflected in higher NJD than some competing methods, and it has increased model complexity and GPU memory footprint due to repeated FIFM and attention layers. Future work is described as the incorporation of biomechanical or adversarial constraints for stricter diffeomorphism and model pruning or distillation for deployment on resource-limited hardware.

The clinical implications identified in the paper follow directly from the evaluated scenarios and the unsupervised design. LGANet++ is presented as suitable for atlas-based segmentation, longitudinal disease monitoring such as tumor growth or atrophy, intraoperative guidance such as MRI–CT fusion, and multi-modal diagnostic fusion such as PET–MRI. Its reported inference time of less than 1 second per volume further suggests utility in time-sensitive workflows without requiring manual annotations [2601.14337].

Source: https://www.emergentmind.com/topics/lganet