BMVC 2026 Accepted paper

GatedSRP

Gated Spatial Redundancy Projection for Pathology Transformer Attentions

A lightweight attention correction that adapts each patch-token update to the tissue context around it.

Zhiyuan Yang · Jiahao Cheng · Vincent Quoc-Huy Trinh · Mahdi S. Hosseini

Paper Code

          ~0.02%
          additional trainable parameters
        

          +0.0269
          mean C-index gain over base attention across five TCGA survival cohorts
        

GatedSRP pipeline from spatial patch neighborhoods through signed gated redundancy projection and slide-level prediction

How GatedSRP works

Nearby WSI patches often repeat the same tissue pattern. GatedSRP separates that shared local component from the attention update, then learns whether it should be reduced, retained, or reinforced.

Whole-slide examples showing spatially varying local redundancy across tissue regions — Local feature similarity varies within and across slides, motivating a context-dependent intervention.

Local reference. For each patch token, GatedSRP averages the detached value vectors in its valid 3x3 neighborhood and normalizes the result into a local tissue-redundancy axis, r̂ᵢ = normalize(mean_j∈N(i) vⱼ).

Adaptive signed correction. A token- and head-specific gate predicts βᵢ = δ tanh(gᵢ) and updates the attention contribution as zᵢ = yᵢ − βᵢ ⟨yᵢ, r̂ᵢ⟩ r̂ᵢ. A negative β reinforces useful local context, β = 0 preserves the original update, 0 < β ≤ 1 reduces the redundant component, and 1 < β ≤ 2 reflects it beyond full projection. Zero initialization starts from the preserve case.

Results

Five-seed evaluation across survival, classification, MIL models, design choices, and measured resource use.

TCGA overall survival

Case-level C-index, mean across five seeds. GatedSRP improves over the no-adaptation baseline on all five cohorts; the across-cohort paired estimate is +0.0269, 95% CI [0.0148, 0.0389], p = 0.0035.

Green indicates an increase and red a decrease relative to NA.

Method	KIRC	KIRP	LUAD	STAD	UCEC
NA	0.7110	0.7247	0.5513	0.5910	0.6756
XSA	0.7192	0.6795	0.5359	0.6035	0.6233
Diff	0.7241	0.7020	0.5546	0.5847	0.6489
GatedSRP	0.7257	0.7648	0.5832	0.6171	0.6973

WSI classification

Selected task metric, mean across five seeds. GatedSRP improves over the no-adaptation baseline on four of five datasets and on 12 of 16 reported classification metrics.

Green and red follow the meaningful changes reported relative to NA in the paper.

Method	CAM16 F1	CAM17 F1	KGH F1	PANDA kappa	BRACS F1
NA	.9731	.7836	.8297	.8822	.4179
XSA	.9730	.7342	.8469	.8811	.4328
Diff	.9923	.7578	.8471	.8826	.3874
GatedSRP	.9808	.7785	.8544	.8842	.4424

The selected metric improves over NA on four datasets; CAMELYON17 is the exception. Per-seed values and all task metrics are included in the repository.

MIL survival models

Case-level C-index across selected public TCGA cohorts.

Method	KIRP	LUAD	STAD	UCEC
ABMIL	.7224	.5553	.5905	.6529
DSMIL	.7253	.5780	.5833	.6621
Official TransMIL	.7318	.5422	.5856	.6460
GatedSRP	.7648	.5832	.6171	.6973

Neighborhood size

The 3x3 neighborhood is best on all six evaluated datasets.

Dataset	3x3	5x5	7x7
KIRP	.7648	.7368	.7275
LUAD	.5832	.5633	.5763
STAD	.6171	.5861	.5694
KGH	.8544	.8429	.8312
PANDA	.8842	.8720	.8787
BRACS	.4424	.4209	.4210

Coefficient parameterization

Fixed signed range versus direct soft-clipped β = 2 tanh(g/2).

Dataset	Fixed	Direct β
KIRP	.7648	.7582
LUAD	.5832	.5619
STAD	.6171	.5732
KGH	.8544	.8407
PANDA	.8842	.8870
BRACS	.4424	.4239

Resource profile

Three-epoch paired profiles. Throughput is in whole slides per second and memory is peak reserved GPU GiB.

Method	PANDA memory	PANDA train	PANDA test	TCGA-5 mean memory	TCGA-5 mean train
NA	.47	11.806	46.588	1.80	1.910
XSA	.52	12.081	53.365	1.92	1.903
Diff	.57	7.680	33.493	2.08	1.896
GatedSRP	.49	9.790	32.962	4.69	1.925

The streaming correction limits PANDA memory, while variable-length TCGA bags produce a higher peak reserved-memory profile.

Learned gate behavior

Most token-level means lie between zero and one, but the model can preserve the original update, apply weak negative correction, or move beyond full projection when the context supports it.

Learned gate coefficient distributions across classification and survival datasets — Learned coefficient distributions vary by dataset and slide rather than collapsing to one fixed value.

Examples of weak negative, identity-like, projection, and amplification coefficient regimes — Representative signed-gate regimes. None of 410 evaluated checkpoint exports had a mean coefficient above 1.5.

Attention heatmaps comparing baseline and GatedSRP across pathology slides — Attention remains slide-specific while the local correction changes how redundant evidence contributes to the representation.

How to use

Add one correction between the patch-token attention update and the residual connection. The attention operator, positional encoding, CLS token, and task head can remain unchanged.

Your model needs three things

Patch coordinatesBuild a 3x3 graph aligned with the patch-token order.
Patch attention updatesCorrect patch tokens only; special tokens pass through unchanged.
A pre-residual insertion pointApply GatedSRP after attention and train normally from its identity initialization.

Minimal integration

import torch

from slide_level_srp.src.srp_correction import PatchSRPCorrection

self.gated_srp = PatchSRPCorrection(
    dim, hidden_dim=32, delta_scale=2.0,
)

update = self.attn(self.norm(x))
patches = self.gated_srp(
    update[:, 1:],
    neighbor_index,
    neighbor_mask,
)
update = torch.cat((update[:, :1], patches), dim=1)
return x + update

Citation

BibTeX

@misc{yang2026gatedspatialredundancyprojection,
  title={Gated Spatial Redundancy Projection for Pathology Transformer Attentions},
  author={Zhiyuan Yang and Jiahao Cheng and Vincent Quoc-Huy Trinh and Mahdi S. Hosseini},
  year={2026},
  eprint={2608.08374},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2608.08374}
}