Symmetry Interactive Transformer for AD Diagnosis

• SIT is a symmetry-aware transformer that explicitly compares paired hemispheric features from sMRI to detect Alzheimer’s-induced asymmetry.
• The model integrates a 3D CNN encoder with cross-hemispheric attention, using a HemiSim Matrix to fuse left–right features effectively.
• Empirical evaluations on ADNI data show that proper hemispheric alignment and symmetry-aware mechanisms significantly boost diagnostic accuracy.

Searching arXiv for papers on “Symmetry Interactive Transformer” and closely related transformer/symmetry formulations. Symmetry Interactive Transformer (SIT) most directly denotes a symmetry-aware transformer module introduced within an end-to-end 3D CNN–transformer framework for Alzheimer’s disease (AD) diagnosis from structural MRI (sMRI), where the defining objective is to model asymmetric structural degeneration between the left and right cerebral hemispheres (Yang et al., 10 Sep 2025). In that formulation, hemispheres are explicitly separated, anatomically aligned, encoded by a 3D CNN, and compared through cross-hemispheric attention biased by a learned symmetry signal. The acronym SIT is not unique across an unrelated graph-theoretic usage denotes “summable intersection tail” (Perl, 2018), and adjacent transformer literature includes both a conceptual symmetry-reduction framework for relational transformers (François et al., 21 Feb 2026) and a distinct “Symmetry-Invariant Transformer” (SiT) for reinforcement learning (Weissenbacher et al., 2024). In current transformer research, however, “Symmetry Interactive Transformer” most specifically refers to the hemispheric interaction architecture for sMRI-based AD diagnosis.

1. Nomenclature and conceptual scope

The term SIT is polysemous across the literature, and disambiguation is necessary because the same acronym names technically unrelated objects.

Usage	Paper	Meaning
SIT	(Yang et al., 10 Sep 2025)	Symmetry Interactive Transformer for AD diagnosis from sMRI
SiT	(Weissenbacher et al., 2024)	Symmetry-Invariant Transformer for generalisation in reinforcement learning
SIT-like	(François et al., 21 Feb 2026)	A symmetry-reduced, relational transformer framework; the paper does not use the exact name
SIT	(Perl, 2018)	Summable intersection tail in graph transience theory

Within machine learning, the AD-diagnosis SIT and the symmetry-reduction framework share a family resemblance only at a high level: both treat symmetry as structurally significant rather than incidental. The former uses anatomical left–right correspondence and disease-induced asymmetry as an explicit diagnostic prior (Yang et al., 10 Sep 2025). The latter argues that transformer computation and optimization should be reformulated in terms of invariant relational quantities so that redundant coordinate choices are removed by construction (François et al., 21 Feb 2026). A plausible implication is that these works occupy different points on the same spectrum of symmetry-aware design: one encodes domain symmetry into attention interactions, while the other seeks quotient-space formulations of transformer representations and learning dynamics.

The graph-theoretic SIT is entirely unrelated. There, SIT stands for summable intersection tail, a property of path measures on graphs equivalent to transience in the edge-intersection sense (Perl, 2018). That usage has no connection to transformers or neural architectures.

2. Pathological motivation: asymmetry as a diagnostic signal

The SIT architecture for AD diagnosis is motivated by a clinical observation emphasized in the paper: AD does not merely induce generic atrophy, but often produces asymmetric structural degeneration between the two hemispheres (Yang et al., 10 Sep 2025). The model therefore treats hemispheric asymmetry as a first-class biomarker rather than as an incidental byproduct of whole-brain processing.

The pipeline begins by segmenting the input sMRI volume along the sagittal plane into left and right hemispheres. The right hemisphere is then flipped so that it is orientation-aligned with the left hemisphere. The image is subsequently divided into isometric 3D grid blocks such that each block on one side has a symmetric counterpart on the other. The authors describe this pairing step as the inter-equal grid block fetch operation (Yang et al., 10 Sep 2025).

This design establishes a strict notion of contralateral correspondence. The network is not comparing arbitrary volumetric regions; it is comparing paired anatomical locations. The paper argues that many existing deep-learning approaches either process the brain as a single undifferentiated volume or use hemisphere separation only as preprocessing rather than as a learnable interaction mechanism. SIT instead makes left–right comparison internal to the model.

A central interpretive consequence follows. High similarity between paired hemispheric features suggests normal symmetry, whereas reduced similarity can indicate disease-driven asymmetry. The architecture is thus organized around detecting departures from bilateral correspondence rather than only identifying absolute local abnormalities. This suggests a diagnostic regime in which asymmetry is treated as a structured relational feature, not merely an intensity pattern.

3. Architecture and mathematical formulation

The complete framework has four components: input block fetching / hemisphere splitting, patch-level 3D CNN encoder, Symmetry Interactive Transformer (SIT), and classifier (Yang et al., 10 Sep 2025).

The 3D CNN encoder processes each corresponding left/right patch to extract local volumetric features. It contains four convolutional modules, each composed of 3D convolution, batch normalization, and ReLU. The first convolution uses a $4 \times 4 \times 4$ kernel, and the next three use $3 \times 3 \times 3$ kernels. If the input patch tensor has shape $[B \cdot N, 1, H, W, D]$, the encoder outputs feature maps of shape $[B \cdot N, C, H, W, D]$, where $C = 128$. After flattening, patch features are represented as sequences with shape $[B, N, C]$ (Yang et al., 10 Sep 2025).

SIT itself is built from multiple Symmetry Interactive Units (SIUs). Its defining mechanism is cross-hemispheric interaction: the query is taken from one hemisphere, while keys and values are taken from the opposite hemisphere. This differs from ordinary within-stream self-attention because it forces the model to compare contralateral regions directly. The paper describes this as being inspired by cross-attention but augmented with an explicit symmetry prior.

That prior is encoded by the HemiSim Matrix (HSM), which measures similarity between corresponding left and right hemisphere features using cosine similarity:

$$HSM = \frac{\sum_{i=1}^{n} \phi_i \psi_i}{\sqrt{\sum_{i=1}^{n} \phi_i^2} \cdot \sqrt{\sum_{i=1}^{n} \psi_i^2}}.$$

Here, $\phi_i$ denotes the $i$-th feature from the left hemisphere and $\psi_i$ the corresponding feature from the right hemisphere. For input feature shape $3 \times 3 \times 3$0, the HSM has shape $3 \times 3 \times 3$1 (Yang et al., 10 Sep 2025).

The HSM is injected into attention through HemiFuse Attention:

$3 \times 3 \times 3$2

where $3 \times 3 \times 3$3 is the query from one hemisphere, $3 \times 3 \times 3$4 and $3 \times 3 \times 3$5 are the key and value from the opposite hemisphere, $3 \times 3 \times 3$6 is the head dimension, and $3 \times 3 \times 3$7 are learnable scalar weights balancing dot-product attention and the symmetry prior (Yang et al., 10 Sep 2025).

The transformer block follows a residual formulation:

$3 \times 3 \times 3$8

and

$3 \times 3 \times 3$9

with $[B \cdot N, 1, H, W, D]$0 the input token feature, $[B \cdot N, 1, H, W, D]$1 layer normalization, and $[B \cdot N, 1, H, W, D]$2 a feed-forward network (Yang et al., 10 Sep 2025). The model uses multi-head attention so that distinct heads can focus on different forms of inter-hemispheric relation.

After SIT, the token features $[B \cdot N, 1, H, W, D]$3 are passed to a classifier, written in the paper as

$[B \cdot N, 1, H, W, D]$4

Training uses a joint loss combining left-hemisphere, right-hemisphere, and fused-output supervision:

$[B \cdot N, 1, H, W, D]$5

where $[B \cdot N, 1, H, W, D]$6 and $[B \cdot N, 1, H, W, D]$7 are hemisphere-specific cross-entropy terms, $[B \cdot N, 1, H, W, D]$8 is the final fused cross-entropy term, and $[B \cdot N, 1, H, W, D]$9 is reported to work well (Yang et al., 10 Sep 2025).

4. Experimental protocol and implementation details

The experiments are conducted on the ADNI dataset, comprising 801 sMRI images: 213 AD, 234 CN, 219 sMCI, and 135 pMCI (Yang et al., 10 Sep 2025). The split is 70% training, 15% validation, and 15% test.

Preprocessing includes correction steps on a clinical platform, affine alignment to MNI152, and skull stripping. The final preprocessed image resolution is $[B \cdot N, C, H, W, D]$0 (Yang et al., 10 Sep 2025). These steps establish a consistent anatomical frame before the hemisphere-specific operations are applied.

Training is performed with Adam on an NVIDIA RTX 4060Ti-16GB GPU, using batch size 4 for 70 epochs. The learning-rate schedule is piecewise: $[B \cdot N, C, H, W, D]$1 for the first 30 epochs, $[B \cdot N, C, H, W, D]$2 for epochs 30–50, and $[B \cdot N, C, H, W, D]$3 thereafter (Yang et al., 10 Sep 2025).

The evaluation metrics are ACC, SEN, SPE, and AUC. The paper also reports hyperparameter findings. The best result is obtained with patch size 25, yielding 92.54% ACC, 92.45% AUC, and 3.77M parameters. The preferred transformer configuration uses SI Units = 5 and Heads = 4 (Yang et al., 10 Sep 2025).

These details are important because the architecture’s efficacy depends on both symmetry-aware attention and correct anatomical alignment. The role of the flipped hemisphere and the inter-equal grid block fetch operation is therefore not incidental preprocessing but part of the model’s operational definition.

5. Empirical performance and ablation structure

The headline result on AD vs CN is achieved by LRCNN-SIT(FLIP), with 92.54% ACC, 87.50% SEN, 97.14% SPE, and 92.32% AUC (Yang et al., 10 Sep 2025). On the more difficult pMCI vs sMCI task, the model achieves 77.36% ACC, 75.00% SEN, 78.79% SPE, and 76.89% AUC.

The ablation study isolates the contribution of hemispheric decomposition, interaction, symmetry-aware attention, and flipping-based alignment.

Variant	AD vs CN ACC	pMCI vs sMCI ACC
wCNN	80.60%	69.81%
LRCNN	86.57%	—
LRCNN-T	86.57%	75.47%
LRCNN-IT	88.06%	—
LRCNN-SIT	85.07%	—
LRCNN-SIT(FLIP)	92.54%	77.36%

Several patterns are explicit in these results (Yang et al., 10 Sep 2025). First, hemisphere splitting alone is useful: LRCNN improves substantially over the whole-brain wCNN on AD vs CN. Second, a generic transformer does not materially improve AD vs CN over LRCNN, although LRCNN-T improves the harder pMCI vs sMCI task. Third, cross-hemispheric interaction is beneficial, as shown by LRCNN-IT. Fourth, LRCNN-SIT without flipping performs worse than the full model, indicating that symmetry-aware attention is not sufficient if contralateral correspondence is misaligned.

The strongest ablation conclusion is that both alignment and symmetry-aware attention are necessary. The paper states that the HSM alone is not enough if corresponding patches are misregistered. This is the central empirical justification for the “interactive” designation in SIT: the model does not simply aggregate symmetric features; it learns from structured hemispheric interaction under correct spatial pairing (Yang et al., 10 Sep 2025).

The proposed method is also reported to outperform several recent AD diagnosis baselines on ADNI, including CNNViT, MTFIL-Net, ATTCNN, AGHyb-Net, DA-MIDL, and AAGN. The paper emphasizes especially the high specificity of the full model (Yang et al., 10 Sep 2025).

6. Interpretation and relation to broader symmetry-aware transformer research

Interpretability is examined using 3D Grad-CAM applied to the last layer of SIT. The resulting heatmaps indicate attention to atrophic regions, especially asymmetric pathological regions, with the hippocampus highlighted as an example of a structure affected early in AD and MCI (Yang et al., 10 Sep 2025). The authors interpret this as evidence that the network is focusing on clinically meaningful asymmetry rather than arbitrary volumetric patterns.

Within the broader transformer literature, SIT occupies a specific niche. The AD model uses symmetry in the anatomical sense of bilateral correspondence and seeks pathology through departures from that symmetry. By contrast, the relational transformer framework of “Toward Manifest Relationality in Transformers via Symmetry Reduction” reformulates transformers in terms of invariant relational quantities such as Gram matrices and bilinear parameter products, with the aim of eliminating redundant degrees of freedom in model space and head space (François et al., 21 Feb 2026). That work is explicit that it is conceptual and mathematical rather than empirical, and it proposes optimization on reduced spaces or via projection orthogonally to symmetry orbits. A plausible implication is that it addresses symmetry of representation and parameterization, whereas SIT for AD addresses symmetry of biological structure and cross-hemispheric comparison.

A second neighboring line is the Symmetry-Invariant Transformer (SiT) for reinforcement learning, presented as a scalable vision transformer that leverages local and global data patterns in a self-supervised manner and refines self-attention through Graph Symmetric Attention to preserve graph symmetries, yielding invariant and equivariant latent representations (Weissenbacher et al., 2024). That model is reported to show superior generalization over ViTs on MiniGrid and Procgen, and improved sample efficiency on Atari 100k and CIFAR10 (Weissenbacher et al., 2024). Although terminologically close, it is distinct from the AD-diagnosis SIT in both domain and symmetry target.

Taken together, these works indicate that “symmetry-aware transformer” can mean at least three different things in current arXiv usage: exploiting anatomical left–right asymmetry for medical diagnosis, enforcing invariance or equivariance in structured RL settings, and reducing continuous representational symmetries through quotient-style formulations. SIT in the strict sense of Symmetry Interactive Transformer refers to the first of these: a cross-hemispheric attention mechanism designed to detect disease-based asymmetry from sMRI (Yang et al., 10 Sep 2025).