acceptodds
Under review as a conference paper at ICLR 2027

Beyond Cartesian Scanning: Unifying Conflict and Geometry in State Modeling for Fine-Grained Medical Image Classification

Abstract

Fine-grained medical image classification depends on subtle local variations whose diagnostic significance is closely related to surrounding anatomical context. Effective recognition therefore requires refining weak discriminative features during contextual integration and organizing surrounding evidence through meaningful structural relationships. Although Vision Mamba efficiently captures long-range dependencies, global connectivity alone does not explicitly address these requirements. We propose Conflict Recalibration and Geometric Aggregation (CRGA), a framework that enhances pretrained VMamba through directional-response-guided feature refinement and circular–radial context aggregation. Specifically, Conflict-Conditioned Recalibration (CCR) compares spatially aligned horizontal and vertical scan responses before fusion and uses their discrepancies as conditioning information for subsequent representation refinement. Through a one-hop connection, this signal modulates selective state writing and readout in the next stage, enabling contextual integration to account explicitly for relationships between directional responses. To organize surrounding structural context, Geometric Circular–Radial Structured Aggregation (GCRSA) introduces a geometric prior inspired by angular–radial formulations in medical imaging. It organizes existing tokens into closed circumferential and open radial sequences around the feature-grid center, distinguishing contextual interactions within radial bands from those across radial distances. Both components retain the original feature grid and Cartesian scan order, while zero-initialized modulation projections and residual gates preserve the pretrained backbone mapping at initialization. Experiments on a clinically acquired EUS dataset show that CRGA improves VMamba accuracy by 2.2 percentage points, while evaluations on two public medical image benchmarks demonstrate competitive performance. Same-backbone ablations on EUS show accuracy gains from each component, with the combined model achieving the highest accuracy and F1 among the evaluated variants.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
State Space Models, Fine-Grained Visual Recognition, Medical Image Classification, Anatomical Structure Modeling, Conflict-Aware Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_wgaqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Conflict Recalibration and Geometric Aggregation (CRGA) for medical image classification using a pretrained VMamba backbone. The framework contains two components, and both components preserve the original Cartesian selective-scan paths and recover the pretrained backbone mapping at initialization through zero-initialized projections or residual gates.

Critical strengths

CCR intervenes directly in selective state writing and readout rather than simply appending a generic feature-attention block. Comparing directional responses before Cross-Merge and using the resulting information in the next stage is a clear architectural hypothesis. The spatial alignment and directional unfolding operations are explicitly described.

Critical weaknesses

CCR intervenes directly in selective state writing and readout rather than simply appending a generic feature-attention block. Comparing directional responses before Cross-Merge and using the resulting information in the next stage is a clear architectural hypothesis. The spatial alignment and directional unfolding operations are explicitly described.

The central hypothesis is that cross-axis discrepancies provide useful conditioning information beyond the features already available to VMamba. However, the current ablation primarily establishes that adding CCR improves performance; it does not establish why. In particular, comparing a full conflict map with only the eight lowest-scoring positions simultaneously changes spatial coverage, signal density, and score selection. This does not isolate the usefulness of high-discrepancy regions or the specificity of directional conflict.

Ablations of (B)-only, (C)-only, and joint modulation are missing, as are comparisons of same-stage versus next-stage conditioning. These experiments would help establish whether the proposed intervention point is essential.

GCRSA uses the feature-grid center, which the authors explicitly distinguish from an anatomical or acquisition center. This distinction is important: grouping tokens by radius around the image center does not necessarily group them by anatomical layer, lesion boundary, or invasion depth.

The main EUS improvement is 2.2 percentage points on 334 test images. Five development-fold models are evaluated on the same held-out test set, but only mean results are reported. These runs do not constitute five independent test cohorts.

Table 2 does not include the unmodified VMamba-Small backbone or the CCR-only and GCRSA-only variants. Consequently, the public results cannot establish whether the proposed modules improve over their own backbone outside EUS.

Questions for authors

Does CCR outperform parameter-matched conditioning derived from merged features, constant maps, or spatially shuffled conflict maps?

Why is the single-channel, rank-one modulation sufficient? What happens with (B)-only, (C)-only, direction-specific, or direct feature modulation?

Can the authors provide per-run results and patient-clustered confidence intervals for the differences between VMamba, CCR, GCRSA, and the combined model?

What are the exact grouping parameters and aggregation implementation? What are the measured latency, memory, and resolution-scaling costs?

3 comments
AuthorsOpenReview ↗

Response to Reviewer R-wgaq

We appreciate the reviewer's recognition of the architectural hypothesis behind CCR and the preservation of pretrained VMamba computation. The concerns focus on whether directional discrepancy contributes beyond generic modulation, the role and capacity of the one-channel B/C correction, the interpretation of GCRSA's geometric reference, public-data attribution, and reliability and computational cost. We address these issues with the completed targeted controls and explicit implementation details.

R-wgaq–1. Why a single-channel CCR map, and how are direction and state parameters handled?

The single-channel map is a low-capacity spatial conditioning signal, not a replacement for VMamba's original input- and direction-dependent state parameters. At each spatial location uu, its learned additive corrections can be viewed schematically as ΔB(u)=WBM(u)\Delta B(u)=W_BM(u) and ΔC(u)=WCM(u)\Delta C(u)=W_CM(u). The original state parameters and recurrences remain intact; the low-rank correction limits the additional capacity and uses zero-initialized projections to recover pretrained computation at initialization. We do not claim one channel is the globally optimal capacity.

With aligned horizontal and vertical responses HH and VV, standard additive Cross-Merge F=H+VF=H+V does not uniquely determine the cross-directional relation: (1,1)(1,1) and (2,0)(2,0) have the same sum and different discrepancies. CCR therefore extracts absolute, normalized, and cosine differences before fusion, rather than claiming to recover diagnostic saliency from the merged representation. The map describes difference between directional contextual responses, not which scanning direction is clinically superior or which pixels belong to a lesion.

For a directional scan dd, the implemented conditioning is

\widehat B^{\\,d}&=B^d+\mathcal T_d(P_B(M)),\\\\ \widehat C^{\\,d}&=C^d+\mathcal T_d(P_C(M)). \end{aligned}$$ Here $\mathcal T_d$ aligns the common spatial conditioning map with each directional Cross-Scan sequence. The additional projections $P_B$ and $P_C$ are **shared**, while VMamba's original directional state parameters and selective recurrences remain **direction-specific**. This is **direction-aligned conditioning**, while the original directional state parameters and selective recurrences remain direction-specific. **Table R-wgaq-1. EUS conditioning-source, target, and stage controls.** | Variant | ACC (%) | Recall (%) | F1 (%) | AUC (%) | |---|---|---|---|---| | **Original CCR: next-stage B+C** | 81.3 ± 0.52 | 81.5 ± 0.80 | 82.7 ± 0.55 | 93.5 ± 0.57 | | Feature-BC | 79.0 ± 0.44 | 78.0 ± 3.14 | 79.6 ± 1.82 | 94.0 ± 0.63 | | Constant-BC | 80.0 ± 0.64 | 79.3 ± 1.10 | 80.2 ± 1.35 | 93.8 ± 0.69 | | Shuffled-BC | 79.3 ± 0.45 | 76.3 ± 1.45 | 79.0 ± 1.01 | 92.8 ± 0.52 | | Conflict-Feature | 79.9 ± 0.31 | 76.9 ± 0.58 | 79.3 ± 0.40 | 93.9 ± 0.25 | | B-only | 80.1 ± 0.43 | 78.1 ± 1.72 | 80.3 ± 1.38 | 93.3 ± 0.51 | | C-only | 80.3 ± 0.90 | 79.4 ± 1.29 | 80.7 ± 1.27 | 94.0 ± 0.64 | | Same-stage (subsequent block; no feedback) | 78.9 ± 0.68 | 77.8 ± 1.23 | 79.3 ± 0.81 | 93.3 ± 0.40 | *Values are percentages. ± denotes run-level SD.* Feature-BC versus original CCR tests **conditioning information with the same B/C target**; Conflict-Feature tests **direct feature versus B/C intervention with the same discrepancy source**. Constant and Shuffled maps test spatially varying content and alignment. B-only, C-only, and Same-stage further isolate state writing, readout, and intervention location. Original CCR gives the highest observed ACC among these controls. The experiments support the particular combination of **cross-directional information and next-stage B/C conditioning**.

R-wgaq–2. Matched selection and interpretation of conflict scores

The original Full-map versus Low-8 ablation changes both coverage and selection. We added Top-8 and Random-8 controls, keeping eight positions at each applicable stage in all sparse variants. Top/Low use each input's largest/smallest discrepancy scores; Random-8 uses predetermined score-independent positions held fixed across training and testing.

Table R-wgaq-2. Full and matched sparse CCR selection on EUS.

SelectionACC (%)Recall (%)F1 (%)AUC (%)
Full Conflict Map81.3 ± 0.5281.5 ± 0.8082.7 ± 0.5593.5 ± 0.57
Top-880.4 ± 0.2579.2 ± 0.2881.2 ± 0.6693.3 ± 0.26
Random-878.8 ± 0.2476.9 ± 2.6278.5 ± 1.5693.5 ± 0.17

Values are percentages. ± denotes run-level SD.

Full-map conditioning yields the highest observed ACC, although comparison with sparse maps still changes coverage. Under matched budgets, Low-8 is slightly above Top-8, while both exceed Random-8; Random-8 uses fixed score-independent positions. The results support using conflict scores as a spatially distributed relational signal for downstream state computation. This interpretation is reinforced by the Feature-BC, Constant-BC, and Shuffled-BC controls in R-wgaq-1.

R-wgaq–3. GCRSA's center, exact nature of aggregation, and geometric controls

Preprocessing and center definition. EUS images were manually cropped to retain the complete esophageal structure. The crop boundaries were determined by organ completeness and did not use probe location. We did not perform probe detection, probe-centered alignment, or anatomical-center estimation. Each cropped image was directly resized to the target resolution without padding. The GCRSA reference is therefore the center of the resized image grid and is not guaranteed to coincide with the transducer or an anatomical center. We use it as a deterministic computational reference. Circular and radial groups encode complementary geometric token neighborhoods around this reference.

Existing tokens within the inscribed circle are ordered into concentric ring sequences, and grid tokens are organized into angular sectors for center-to-periphery radial sequences. Clockwise/counter-clockwise and inward/outward propagation use learned relative-offset weights shared across channels, followed by zero-initialized residual fusion. The original Cartesian selective-scan paths, spatial grid, and token count are preserved. Mathematically, this is an input-independent offset-weighted aggregation within a learned residual branch, not input-adaptive attention or a claim to have invented linear filtering in polar order.

The exact stage-wise grouping and profile configurations are reported in Table G3.

Table R-wgaq-3. EUS geometric organization and alternative implementations.

VariantACC (%)Recall (%)F1 (%)AUC (%)
VMamba-Small79.9 ± 0.5776.5 ± 2.3079.9 ± 1.0094.1 ± 0.70
Full GCRSA81.1 ± 0.4181.2 ± 0.1081.6 ± 0.8595.0 ± 0.61
Uniform Ring/Ray77.7 ± 1.0976.1 ± 1.5278.1 ± 1.7293.3 ± 0.47
Polar-transformed Input75.6 ± 1.1674.1 ± 2.1176.8 ± 1.9891.1 ± 0.73
Feature-level Polar-SSM73.5 ± 2.2771.1 ± 5.0972.7 ± 3.9290.8 ± 0.76

Values are percentages. ± denotes run-level SD.

Uniform Ring/Ray and the polar alternatives perform below Full GCRSA in this comparison. The results favor retaining pretrained Cartesian scanning and adding geometric relationships residually.

The geometry was also perturbed directly: four 5% displacements of the computational center give 81.0% mean ACC (−0.1 pp); four 5% input image translations give 80.8% (−0.3 pp) compared with unperturbed GCRSA (81.1%). The image translation affects the whole backbone; the results reflect bounded average sensitivity, within the evaluated perturbation range.

For the measured inference implementation, Eq. (13) is evaluated using cached geometric indices, reused circular transformation intermediates with separate CW/CCW weights, and parallel masked bidirectional radial accumulation and normalization. Valid offsets and boundary conditions are maintained without polar resampling or extra tokens; the exact stage-specific configurations are given in Table G3.

R-wgaq–4. Public-benchmark competitiveness and component attribution

The two public benchmarks primarily test Full CRGA's competitiveness and applicability to additional medical classification tasks. We also evaluated CCR-only and GCRSA-only alongside the original backbone to answer the reviewer's separate request for component-level attribution.

Table R-wgaq-4. Public VMamba-family comparisons (ACC, %).

MethodFetal-Planes-DBOrganAMNIST-224
VMamba-Small94.895.2
CCR-only95.496.1
GCRSA-only95.296.2
Full CRGA95.596.3

Full CRGA improves over VMamba-Small by +0.7/+1.1 pp on Fetal-Planes-DB/OrganAMNIST-224. Both modules individually exceed the shared backbone on both datasets: CCR-only by +0.6/+0.9 pp, and GCRSA-only by +0.4/+1.0 pp. Thus, component-level improvements are not restricted to EUS. The complete architecture achieves the highest observed ACC on both public benchmarks. Detailed mechanism tests remain centered on the primary EUS task.

R-wgaq–5. EUS variability, paired patient uncertainty, and computational costs

Per-run test accuracies for all four models are provided in Table R-wgaq-5a.

The main EUS dataset includes 1,658 images from 822 patients, with 1,324 training and 334 test images separated at the patient level. For each training fold, the checkpoint with the best validation performance is selected and evaluated on the same patient-separated test cohort, so the five runs are not five independent test populations. The manuscript ACCs for VMamba-Small, CCR-only, GCRSA-only, and Full CRGA are 79.9%, 81.3%, 81.1%, and 82.1%, with sample SDs across the five runs of 0.57, 0.52, 0.41, and 0.16 pp, respectively.

Table R-wgaq-5a. Per-run EUS test accuracy (ACC, %).

ModelRun 1Run 2Run 3Run 4Run 5Mean ± SD
VMamba-S (VSSM)79.079.879.980.480.479.9 ± 0.57
CCR-only82.181.381.181.180.781.3 ± 0.52
GCRSA-only81.481.480.780.681.481.1 ± 0.41
Full CRGA82.182.382.281.982.082.1 ± 0.16

Each run corresponds to a model trained on one development fold and evaluated on the same patient-separated held-out test cohort. SD is the sample standard deviation across the five reported runs (ddof = 1).

Table R-wgaq-5b. Selected patient-level paired EUS differences, exact McNemar p-values, and patient-level 95% CIs.

ComparisonΔACC (pp)Exact McNemar p-valuePatient-level 95% CI (pp)
Full CRGA − VMamba-Small+2.20.076813[−0.24, +4.63]
Full CRGA − CCR-only+0.80.035156[+0.06, +1.62]
Full CRGA − GCRSA-only+1.000.026604[+0.12, +1.88]

Full CRGA exceeds both individual components, with exact McNemar p-values of 0.035156 versus CCR-only and 0.026604 versus GCRSA-only. The Full CRGA versus VMamba-Small comparison shows an observed +2.2 pp gain, with p=0.076813, close to the 0.05 threshold.

Table R-wgaq-5c. Inference costs at 224 × 224, batch size 1.

MethodParameters (M)Latency (ms)Peak allocated memory (MB)
VMamba-Small49.49.1223.0
CCR-only49.49.7223.1
GCRSA-only49.513.2225.1
Full CRGA49.513.7225.1

Full CRGA adds approximately 0.1 M parameters, 4.6 ms (+50.5%) model-forward latency, and 2.1 MB peak allocated GPU memory at 224 × 224. At 384 × 384, the added latency is 9.4 ms (+74.6%) in the evaluated accelerated inference setup. These measured costs remain within a practical range for the intended computer-aided clinical workflow, where the model is designed to assist physicians.

Reviewer_SftiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes CRGA, a VMamba-based framework for fine-grained medical image classification. CCR uses disagreement between directional scan responses to modulate state writing and readout in the next stage, while GCRSA aggregates features along circular and radial groups. The method is evaluated on an in-house EUS dataset and two public benchmarks, with component and design comparisons on EUS.

Critical strengths

Controlled evaluation on the main dataset. The patient-level splits and shared backbone, initialization, and training settings make the EUS comparisons informative. Both modules improve accuracy individually, and their combination raises accuracy from 79.9% to 82.1%, providing encouraging evidence that the additions are useful.

A concrete way to exploit directional information. CCR uses differences between scan responses before they are merged and carries this information into subsequent state computation. This is an interesting design choice, and the circular–radial branch provides a complementary spatial organization with a modest reported FLOPs increase.

Critical weaknesses

The experiments do not sufficiently isolate the proposed mechanisms. Comparing the full conflict map with eight lowest-scoring positions changes both spatial coverage and selection strategy. Without matched-selection controls or a simpler feature-modulation comparison, it remains unclear whether directional conflict and state-parameter modulation specifically explain the gains.

The evidence beyond the main EUS comparison is incomplete. The public-dataset results omit the unmodified VMamba backbone, so they do not directly establish the benefit of CRGA. The EUS results also lack variability estimates across the five development-fold runs, leaving uncertainty about the stability of the relatively small gains.

Questions for authors

The paper explains how CCR and GCRSA work, but their contribution beyond feature-conditioned modulation and structured aggregation could be clearer. In particular, what does conditioning state computation on directional disagreement offer beyond a simpler feature-modulation approach?

The comparison between the full conflict map and eight lowest-scoring positions changes both the selection strategy and spatial coverage. Comparing high-score, low-score, and random positions while keeping the number of selected positions fixed would help isolate the value of the conflict scores.

Using the feature-grid center is a simple geometric choice, but it may not align with the relevant anatomical structure. Tests with shifted centers or translated images would help show how sensitive GCRSA is to this choice.

CCR jointly modulates (B) and (C) in the next stage. Separate results for (B)-only and (C)-only modulation, along with a direct feature-modulation control, would help explain which part of this design drives the improvement.

Table 2 does not include the unmodified VMamba backbone. Please add this baseline under the same training settings on both public datasets so that the gains from CRGA can be assessed directly. Component-level results would also help establish whether both modules remain useful beyond EUS.

The EUS results average five development-fold runs, but their variability is not reported. Please include the spread across runs and confidence intervals for the paired performance differences, taking into account that some test images come from the same patient.

The small FLOPs increase is encouraging. Since geometric aggregation also involves indexing and data movement, inference latency and peak memory would give a more complete picture of its practical cost. A brief explanation of how Eq. (13) is implemented efficiently would also be useful.

4 comments
AuthorsOpenReview ↗

Response to Reviewer R-Sfti

We thank the reviewer for recognizing the improvements of CRGA over the shared VMamba-Small backbone and the observed contributions of CCR and GCRSA on the EUS task. The reviewer also raises detailed questions about whether these gains arise from the proposed directional conditioning and circular–radial organization rather than simpler alternatives. We address these questions in five connected themes covering the conditioning source and intervention, spatial selection and modulation targets, geometric organization, public-benchmark attribution, and statistical and computational evaluation.

R-Sfti–Q1. CCR and GCRSA beyond simpler alternatives

The reviewer asks what conditioning state computation on directional disagreement adds beyond a simpler feature-modulation approach, and whether GCRSA provides more than an additional structured aggregation branch. The distinction is where the proposed signals enter the pretrained computation. Feature modulation changes the representation before selective scanning, whereas CCR preserves the directional discrepancy before Cross-Merge and uses it to condition the next-stage state-writing parameter BB and state-readout parameter CC. The same discrepancy signal therefore participates directly in recurrent state update and readout. We test this distinction by comparing CCR with Feature-BC and Conflict-Feature, which use the same or related discrepancy information but intervene through feature modulation, and with B-only, C-only, and Same-stage controls. For GCRSA, we compare the learned circular–radial organization with simpler aggregation alternatives to determine whether the gain comes from the proposed geometric relationships rather than from adding an arbitrary branch. In VMamba, spatially aligned horizontal and vertical selective-scan responses are additively merged: F=H+VF=H+V. Different directional pairs can yield the same FF with different H−VH-V; the sum therefore does not uniquely retain their relation. CCR extracts absolute, normalized, and cosine discrepancies before Cross-Merge. This is a representation of cross-directional response discrepancy, not a measure of clinical disagreement, lesion saliency, or uncertainty.

The discrepancy map then modifies the existing selective-state input and readout parameters in the next stage's first VSS block. Schematically,

y_t=C_t h_t+D x_t.$$ Conditioning $B$ explicitly changes state writing and may affect subsequent recurrent states; conditioning $C$ changes readout. Feature modulation can also change these quantities indirectly, so the architectural distinction is **explicit state-parameter targeting**, not an exclusive ability to affect history. The added projections are zero-initialized to preserve the pretrained mapping at initialization. **Table R-Sfti-1. EUS controls isolating CCR conditioning source and intervention point.** | Configuration | ACC (%) | Recall (%) | F1 (%) | AUC (%) | |---|---|---|---|---| | **Original CCR (B+C)** | 81.3 ± 0.52 | 81.5 ± 0.80 | 82.7 ± 0.55 | 93.5 ± 0.57 | | Feature-BC | 79.0 ± 0.44 | 78.0 ± 3.14 | 79.6 ± 1.82 | 94.0 ± 0.63 | | Conflict-Feature | 79.9 ± 0.31 | 76.9 ± 0.58 | 79.3 ± 0.40 | 93.9 ± 0.25 | | Constant-BC | 80.0 ± 0.64 | 79.3 ± 1.10 | 80.2 ± 1.35 | 93.8 ± 0.69 | | Shuffled-BC | 79.3 ± 0.45 | 76.3 ± 1.45 | 79.0 ± 1.01 | 92.8 ± 0.52 | *Values are percentages. ± denotes run-level SD.* Feature-BC uses merged-feature statistics to generate a conditioning map with the same B/C projection. Conflict-Feature preserves the original discrepancy map but directly modulates internal features before Cross-Scan. Constant-BC and Shuffled-BC control map variation and spatial correspondence. The **81.3%** original CCR result exceeds all four tested alternatives, providing direct empirical support for the joint choice of **directional discrepancy as the conditioning source and B/C as the intervention target**. This result favors the CCR conditioning source and B/C intervention target among the evaluated alternatives. The GCRSA contribution is complementary. It introduces circular–radial token neighborhoods while preserving pretrained Cartesian selective scans; its value lies in **which relationships are aggregated and how the branch is integrated**, not in inventing a new general filtering operator. In EUS controls, Full GCRSA attains **81.1% ACC**, compared with **77.7%** for Uniform Ring/Ray aggregation. This cross-evidence further indicates that the observed benefit is not explained simply by appending an arbitrary structured aggregation branch.

R-Sfti–Q2. Spatial selection and modulation targets

The original comparison used the complete conflict map as Full and the eight positions with the lowest learned discrepancy scores as Low-8. This comparison was designed to test whether retaining directional conflict information improves the result relative to a sparse low-conflict condition. We then added Top-8, which selects the eight highest-discrepancy positions, and a Random-8 control. Top-8 selection is straightforward because it follows the learned discrepancy scores. Resampling random positions during training caused unstable optimization, so Random-8 uses score-independent positions generated from a predetermined seed and held fixed during training and evaluation. These fixed positions are not selected for their conflict values and may contain little or no directional conflict, providing a practical control for whether an arbitrary sparse selection can reproduce the benefit of conflict-guided selection.

Table R-Sfti-2. Full and equal-budget sparse selection on EUS.

SelectionACC (%)Recall (%)F1 (%)AUC (%)
Full Conflict Map81.3 ± 0.5281.5 ± 0.8082.7 ± 0.5593.5 ± 0.57
Top-880.4 ± 0.2579.2 ± 0.2881.2 ± 0.6693.3 ± 0.26
Random-878.8 ± 0.2476.9 ± 2.6278.5 ± 1.5693.5 ± 0.17

Values are percentages. ± denotes run-level SD.

Full-map conditioning yields the highest observed ACC (81.3%). At matched eight-position budgets, Low-8 (80.5%) and Top-8 (80.4%) both exceed the original VMamba-Small baseline (79.9%), while Random-8 reaches 78.8%. The gains from both Low-8 and Top-8 show that selecting positions according to the learned directional discrepancy provides useful conditioning information, whereas arbitrary fixed positions do not reproduce the same benefit. Top-8 and Low-8 adapt their positions to each input, while Random-8 uses fixed score-independent positions. These results support the effectiveness of the directional conflict signal for subsequent state computation.

We next compare B-only, C-only, joint B+C, and direct feature modulation using the same conflict signal.

CCR specifically conditions the next-stage state-writing parameter BB and state-readout parameter CC. To isolate their roles, we trained B-only and C-only variants under the corresponding setup, alongside the joint B+C design.

Table R-Sfti-4. EUS modulation-target controls.

ConfigurationACC (%)Recall (%)F1 (%)AUC (%)
B-only80.1 ± 0.4378.1 ± 1.7280.3 ± 1.3893.3 ± 0.51
C-only80.3 ± 0.9079.4 ± 1.2980.7 ± 1.2794.0 ± 0.64
Joint B+C (original CCR)81.3 ± 0.5281.5 ± 0.8082.7 ± 0.5593.5 ± 0.57
Conflict-Feature79.9 ± 0.3176.9 ± 0.5879.3 ± 0.4093.9 ± 0.25

Values are percentages. ± denotes run-level SD.

Joint B+C has the highest observed ACC among these alternatives. Conflict-Feature keeps the same discrepancy information but applies direct feature modulation, while B-only/C-only retain state targeting and isolate writing/readout. A separate same-stage intervention reaches 78.9% ACC (run SD 0.68 pp), below the proposed next-stage design. Together, these comparisons identify joint next-stage conditioning of the state-writing and state-readout parameters as the strongest tested intervention, outperforming B-only, C-only, direct feature modulation, and same-stage conditioning.

R-Sfti–Q3. Geometric-center choice and sensitivity to shifts

Preprocessing and center definition. EUS images were manually cropped to retain the complete esophageal structure. The crop boundaries were determined by organ completeness and did not use probe location. We did not perform probe detection, probe-centered alignment, or anatomical-center estimation. Each cropped image was directly resized to the target resolution without padding. The GCRSA reference is therefore the center of the resized image grid and is not guaranteed to coincide with the transducer or an anatomical center. We use it as a deterministic computational reference.

We separately tested 5% reference-center displacement with the image fixed and 5% image translation with the original reference fixed, using four cardinal directions for each intervention. These experiments indicate bounded sensitivity within the tested perturbation range.

The exact stage-wise grouping and profile configurations are provided in Table G3; geometric indices and valid-position masks are cached before inference.

Table R-Sfti-3. Mean EUS ACC across four 5% perturbation directions.

Evaluation settingACC (%)Change (pp)
Original GCRSA81.1—
Reference displacement81.0−0.1
Input image translation80.8−0.3

These means summarize up/down/left/right evaluations on the same trained model and test cohort.

The small average −0.1 pp change indicates that exact alignment to the original grid reference is not essential within the evaluated shift range. Input translation affects the full model, so its −0.3 pp change cannot be assigned solely to GCRSA. The two perturbation tests quantify sensitivity within the evaluated range. The geometric alternatives described in the geometry responses also independently test whether the proposed grouping is preferable to simpler tested organizations.

R-Sfti–Q4. Public same-backbone baselines and component controls

We have added unmodified VMamba-Small, CCR-only, and GCRSA-only on both public benchmarks under the corresponding evaluation settings. These experiments address two distinct issues: whether Full CRGA improves over its own backbone, and whether each component retains positive observed value outside EUS.

Table R-Sfti-5. Public benchmark VMamba-based variants (ACC, %).

MethodFetal-Planes-DBOrganAMNIST-224
VMamba-Small94.895.2
CCR-only95.496.1
GCRSA-only95.296.2
Full CRGA95.596.3

Full CRGA exceeds VMamba-Small by +0.7 pp on Fetal-Planes-DB and +1.1 pp on OrganAMNIST-224. Separately, CCR-only and GCRSA-only improve by +0.6/+0.4 pp on Fetal-Planes-DB and +0.9/+1.0 pp on OrganAMNIST-224. Thus, both components show positive observed gains relative to the shared backbone on both public benchmarks. The complete architecture achieves the highest observed ACC on both public benchmarks. The public datasets principally assess the complete CRGA architecture's competitiveness and wider medical-task applicability, while detailed mechanism attribution is centered on EUS.

R-Sfti–Q5. Statistical variability and computational cost

The five training-fold models are evaluated on the same 334-image patient-separated test cohort. Multiple images may come from the same patient, so variation across trained models and uncertainty of paired test performance differences require different statistical summaries.

Table R-Sfti-6a. EUS run-level summary (ACC and run SD).

MethodACC (%)Recall (%)F1 (%)AUC (%)
VMamba-Small79.9 ± 0.5776.5 ± 2.3079.9 ± 1.0094.1 ± 0.70
CCR-only81.3 ± 0.5281.5 ± 0.8082.7 ± 0.5593.5 ± 0.57
GCRSA-only81.1 ± 0.4181.2 ± 0.1081.6 ± 0.8595.0 ± 0.61
Full CRGA82.1 ± 0.1680.8 ± 0.5582.9 ± 0.2094.4 ± 0.25

Values are percentages. ± denotes run-level SD.

The patient-level analysis reports paired ACC differences, two-sided exact McNemar p-values, and patient-level 95% confidence intervals.

Table R-Sfti-6b. Patient-level paired EUS ACC differences, exact McNemar p-values, and patient-level 95% CIs.

ComparisonΔACC (pp)Exact McNemar p-valuePatient-level 95% CI (pp)
CCR-only − VMamba-Small+1.40.078354[−0.15, +2.87]
GCRSA-only − VMamba-Small+1.200.002350[+0.43, +1.97]
Full CRGA − VMamba-Small+2.20.076813[−0.24, +4.63]
Full CRGA − CCR-only+0.80.035156[+0.06, +1.62]
Full CRGA − GCRSA-only+1.000.026604[+0.12, +1.88]
CCR-only − GCRSA-only+0.160.002599[+0.06, +0.26]

All six observed paired differences are positive. Four of the six comparisons satisfy the two-sided exact McNemar test at p<0.05p<0.05, and the remaining two comparisons have p-values of 0.076813 and 0.078354, both close to the 0.05 threshold.

FLOPs alone do not capture token reordering, index access, and intermediate data movement in geometric aggregation. We therefore measured inference model-forward latency, peak allocated GPU memory, and parameters.

Table R-Sfti-7a. Measured inference costs at 224 × 224, batch size 1.

MethodParameters (M)Latency (ms)Peak allocated memory (MB)
VMamba-Small49.49.1223.0
CCR-only49.49.7223.1
GCRSA-only49.513.2225.1
Full CRGA49.513.7225.1

Full CRGA adds approximately 0.1 M parameters, 4.6 ms of model-forward latency (+50.5%), and approximately 2.1 MB of peak allocated memory compared with VMamba-Small in this configuration. Although CRGA increases measured latency relative to VMamba-Small, the overhead remains within a practical range for the intended computer-aided clinical workflow, where the model is designed to assist physicians.

Table R-Sfti-7b. Additional Full CRGA costs over VMamba-Small, batch size 1.

ResolutionAdded latency (ms)Relative latency increaseAdded peak allocated memory (MB)
224 × 2244.650.5%2.1
256 × 2565.761.3%2.5
320 × 3207.366.5%3.4
384 × 3849.474.6%4.5

These measurements use the evaluated accelerated, frozen-GPU inference implementation. “Peak allocated” is the framework-reported allocated GPU memory, not total device-resident usage. Higher resolutions use shape-adapted offset profiles during evaluation, without additional training.

Efficient implementation of Eq. (13). Geometric grouping indices are precomputed and cached, avoiding repeated spatial-index generation. Circular aggregation reuses intermediate feature transformations for clockwise and counter-clockwise propagation while preserving independent directional weights. Radial aggregation combines bidirectional propagation, valid-position masking, weighted accumulation, normalization, and residual fusion into parallel GPU operations, reducing redundant computation and intermediate data movement. These operations preserve the original geometric relationships, propagation ranges, and boundary conditions without additional tokens or polar resampling.


Reviewer_w5GnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper adds two modules to an ImageNet-pretrained VMamba-Small for fine-grained medical image classification.

  • Conflict-Conditioned Recalibration (CCR) compares the horizontal and vertical (V) selective-scan responses of the last VSS block in a stage. It computes three per-pixel discrepancy descriptors (absolute difference, normalized difference, 1 − cosine similarity) and maps them to a single-channel "conflict" map M_s ∈ [0,1] (Eqs. 6–7). In the next stage, a zero-initialized 1×1 conv projects M_s to additive offsets on the SSM input matrix B and readout matrix C of the first VSS block (Eqs. 8–9).
  • Geometric Circular–Radial Structured Aggregation (GCRSA) groups the Cross-Merged tokens into concentric rings (inside the inscribed circle) and angular sectors around the feature-grid centre. Each ring and sector is aggregated with learned, channel-shared weights over sequence offsets in four directions (clockwise, counter-clockwise, inward, outward), and the result is added back through zero-initialized channel gates (Eqs. 10–14).

On an in-house esophageal EUS dataset (1,658 images, 822 patients, 5 classes, 334-image patient-level held-out test set), CRGA improves VMamba accuracy from 79.9% to 82.1% (Table 1). It reports competitive results on FETAL_PLANES_DB and OrganAMNIST-224 (Table 2), plus module and design ablations on EUS (Tables 3–4).

Critical strengths
  1. A clinically motivated geometric prior. Organizing context into rings and rays around a centre is a natural inductive bias for radial EUS, where tissue layers appear roughly concentric around the transducer.

  2. Implementing it as re-grouping of existing tokens, without polar resampling, is simple. Careful integration with the pretrained backbone. Zero-initialized projections and gates make both modules exact identities at initialization (Secs. 2.2.2, 2.3.2), so fine-tuning starts from the pretrained VMamba.

  3. Sound data handling on EUS. Patient-level splits for both test and cross-validation (Sec. 3.1), and a held-out test set that is never used for model development.

  4. A useful negative control. Replacing circular–radial grouping with horizontal–vertical grouping in the same aggregation form reduces accuracy below the baseline (78.4%; Table 4). This at least suggests the gain is not simply from adding an aggregation branch.

Critical weaknesses
  1. The public benchmark results don't isolate the contribution
  • Table 2 does not include VMamba-S itself on FETAL_PLANES_DB or OrganAMNIST-224, so the improvement attributable to CRGA there is unknown. The comparison is against other architectures, and their numbers appear to be taken from their papers. Training protocols, resolution and augmentation likely differ, and none of this is stated.
  1. The CCR mechanism and its motivation lack controls
  • The conflict map is a single channel, projected to B/C offsets by a 1×1 conv (Eq. 8). The modulation is therefore a single learned vector scaled by a scalar spatial map. It is unclear that the conflict content matters rather than the extra spatially varying bias. Needed controls: (a) a spatially shuffled or random M, (b) a constant M, (c) an M derived from a non-conflict statistic such as ‖H + V‖ or feature norm, with the same projection.
  • The link to inter-rater disagreement (Ji et al., 2021) is only an analogy. There is no argument for why disagreement between horizontal and vertical scan responses should mark diagnostically important regions in fine-grained classification. Fig. 3 shows one qualitative example in which conflict responses are "distributed across the lesion and surrounding tissue," which does not establish this.
  1. GCRSA design clarity
  • Centre choice. The feature-grid centre is the fixed reference (Eq. 10). That is reasonable for radial EUS only if the transducer is centred after preprocessing. Please state whether EUS images are cropped or centred on the probe, and how sensitive results are to off-centre images. Random rotation keeps the centre fixed, but random cropping or translation, if used, would not.
  • What the aggregation actually is. The aggregation is a fixed, input-independent weighted average along each ring or ray with channel-shared weights (Eqs. 12–13), effectively a depthwise linear filter in polar ordering. It would help to compare with simple alternatives that encode the same prior: polar-transformed input or features, polar/spiral scan orders inside the SSM, or rotation-equivariant layers.
  1. Baseline fairness and missing protocol details:
  • Baseline training details (epochs, learning rate, augmentation, checkpoint source, model selection) are deferred to an appendix that is not in the submitted PDF (Sec. 3.1 refers to "the appendix" for dataset statistics and method-specific configurations). It is unclear whether the FGVC and Mamba baselines got any tuning comparable to the 500-epoch VMamba schedule.
Questions for authors
  1. Can you report mean ± SD across the five folds for every table, and a paired significance test (e.g. McNemar on test predictions) of CRGA vs VMamba?
  2. Can you add VMamba-S (same training protocol) to Table 2, and state whether the other Table 2 numbers were retrained or copied?
  3. Can you run CCR controls with a shuffled, constant or non-conflict scalar map using the same B/C projection?
  4. How are EUS images centred and preprocessed relative to the transducer, and how robust is GCRSA to off-centre inputs?

Will raise the score if all my concerns are addressed.

3 comments
AuthorsOpenReview ↗

Response to Reviewer w5Gn

We appreciate the reviewer's constructive assessment, particularly the recognition of GCRSA's geometric motivation, the preservation of pretrained VMamba computation, and the patient-level EUS data partitioning. The requested same-backbone comparisons, CCR mechanism controls, geometric alternatives, and statistical analyses are now available. We address the four numbered questions directly, including the additional concern about baseline provenance and protocol transparency.

R-w5Gn–Q1. Run-level variability, confidence intervals, and paired significance

The EUS study comprises 1,658 images from 822 patients in five categories, with 1,324 training images and 334 patient-separated test images, with no patient appearing in both sets. For each training fold, the checkpoint with the best validation performance is selected and evaluated on the same patient-separated test cohort, not on five independent test populations. Table R-w5Gn-1 reports the manuscript ACC values with the run-level SD.

Table R-w5Gn-1. Main EUS accuracy and run-level variability.

MethodACC (%)Recall (%)F1 (%)AUC (%)
VMamba-Small79.9 ± 0.5776.5 ± 2.3079.9 ± 1.0094.1 ± 0.70
CCR-only81.3 ± 0.5281.5 ± 0.8082.7 ± 0.5593.5 ± 0.57
GCRSA-only81.1 ± 0.4181.2 ± 0.1081.6 ± 0.8595.0 ± 0.61
Full CRGA82.1 ± 0.1680.8 ± 0.5582.9 ± 0.2094.4 ± 0.25

Values are percentages. ± denotes run-level SD.

We performed patient-level paired comparisons using two-sided exact McNemar tests and patient-level confidence intervals. Full CRGA improves over VMamba-Small by +2.2 pp, with p=0.076813p=0.076813 and a 95% CI of [−0.24,+4.63][-0.24,+4.63] pp. CCR-only improves over VMamba-Small by +1.4 pp, with p=0.078354p=0.078354 and a 95% CI of [−0.15,+2.87][-0.15,+2.87] pp. Both comparisons show positive observed gains, and their p-values are close to the 0.05 threshold. The complete set of patient-level paired comparisons is reported in the statistical analysis section.

R-w5Gn–Q2. Public VMamba baselines, sources of Table 2 results, and fairness

Direct same-backbone evidence. We have added unmodified VMamba-Small under the corresponding public benchmark evaluations. This provides the comparison missing from the submitted Table 2.

Table R-w5Gn-2. Complete CRGA versus its own backbone on public benchmarks (ACC, %).

DatasetVMamba-SmallFull CRGAGain (pp)
Fetal-Planes-DB94.895.5+0.7
OrganAMNIST-22495.296.3+1.1

The public datasets primarily assess the competitiveness and broader applicability of Full CRGA, rather than serving as the principal setting for fine-grained mechanism ablation. In both benchmarks, the complete model improves over its unmodified VMamba backbone.

Other entries in Table 2 of the original manuscript come from published comparisons. The ResNet-50, ViT-B, Swin-T, MedViT-S, MedMamba-T, and MedKAFormer-T values were taken from Wang et al., MedKAFormer: When Kolmogorov–Arnold Theorem Meets Vision Transformer for Medical Image Representation ([IEEE JBHI, 2025]). The InceptionMamba reference results were taken from Huang et al., InceptionMamba: A Lightweight and Effective Model for Medical Image Classification Revealing Mamba's Low-Frequency Bias ([Neural Processing Letters, 2026]). We will identify these sources in Table 2 and in the manuscript's experimental setup.

We also independently evaluated representative competing models to provide additional measured reference points:

Table R-w5Gn-3. Independent supplementary evaluations of competing architectures.

DatasetMethodACC (%)Macro AUC (%)
Fetal-Planes-DBInceptionMamba91.798.9
OrganAMNIST-224MedMamba-T94.499.8

When reproduced under the same configurations used for our comparisons, the results of other architectures can differ slightly from their originally reported values. To assess CRGA's comparative competitiveness, the other entries in Table 2 of the original manuscript therefore use the strongest results reported in the corresponding papers. Clearly identifying the source of each literature-reported result is essential, and we will add these source annotations and the associated clarification to the revised manuscript. The independently evaluated VMamba-Small versus CRGA comparisons use the same patient-level partitioning and evaluation protocol described above.

For the principal EUS comparison, all VMamba-based variants use ImageNet-pretrained VMamba-Small, 224 × 224 inputs, AdamW, learning rate 5×10−55\times10^{-5}, weight decay 0.05, batch size 32, a nominal 500-epoch training-fold budget, and random rotation plus horizontal/vertical flips. For each EUS training fold, the checkpoint with the best validation performance is selected before evaluation on the shared patient-separated test cohort. For the public benchmarks, VMamba-Small and CRGA use the same training, augmentation, checkpoint-selection, and evaluation protocol within each dataset, so the reported gains are measured under matched conditions. The revised supplementary material will provide the exact epoch budget, learning-rate schedule, augmentation, checkpoint-selection rule, input resolution, and run count for each public benchmark.

R-w5Gn–Q3. Why directional discrepancy and state-parameter conditioning matter

The reviewer asks whether CCR contributes information beyond a spatially varying feature bias and whether its intervention in the state parameters is necessary. We address this concern through the directional-response construction and the conditioning, spatial, target, and stage controls below.

CCR is not merely an extra spatially varying bias. It captures the relationship between directional responses before they are merged, then uses that signal in subsequent selective-state computation. For spatially aligned bidirectional horizontal and vertical responses,

H&=Y_{h^+}+Y_{h^-},\\\\ V&=Y_{v^+}+Y_{v^-},\\\\ F_{\mathrm{merge}}&=H+V. \end{aligned}$$ The additive fusion is many-to-one: $(H,V)=(1,1)$ and $(2,0)$ produce the same fused scalar response, while $|H-V|$ differs. Therefore, the directional relationship is **not uniquely recoverable from the fused sum alone**. CCR explicitly constructs absolute, normalized, and cosine-based cross-directional discrepancy descriptors before Cross-Merge. This does **not** mean that CCR creates information absent from the full input, nor that high discrepancy is a clinical saliency score. The resulting single-channel map $M$ provides a **low-capacity additive correction** to the existing input-dependent $B/C$ parameters in the first VSS block of the next stage: $$\begin{aligned} \widehat B^{\\,d}&=B^d+\mathcal T_d\\!\left(P_B(M)\right),\\\\ \widehat C^{\\,d}&=C^d+\mathcal T_d\\!\left(P_C(M)\right). \end{aligned}$$ Here $\mathcal T_d$ aligns the condition with the relevant directional Cross-Scan ordering. Conditioning $B$ explicitly affects state writing, whereas conditioning $C$ affects state readout; zero-initialized projections preserve the pretrained computation at initialization. Direct feature modulation may influence state dynamics indirectly, but it does not reproduce this **specific intervention location**. **Table R-w5Gn-4. CCR source, spatial, and intervention controls on EUS.** | Configuration | ACC (%) | Recall (%) | F1 (%) | AUC (%) | |---|---|---|---|---| | **Original CCR (next-stage B+C)** | 81.3 ± 0.52 | 81.5 ± 0.80 | 82.7 ± 0.55 | 93.5 ± 0.57 | | Feature-BC | 79.0 ± 0.44 | 78.0 ± 3.14 | 79.6 ± 1.82 | 94.0 ± 0.63 | | Constant-BC | 80.0 ± 0.64 | 79.3 ± 1.10 | 80.2 ± 1.35 | 93.8 ± 0.69 | | Shuffled-BC | 79.3 ± 0.45 | 76.3 ± 1.45 | 79.0 ± 1.01 | 92.8 ± 0.52 | | Conflict-Feature | 79.9 ± 0.31 | 76.9 ± 0.58 | 79.3 ± 0.40 | 93.9 ± 0.25 | | B-only | 80.1 ± 0.43 | 78.1 ± 1.72 | 80.3 ± 1.38 | 93.3 ± 0.51 | | C-only | 80.3 ± 0.90 | 79.4 ± 1.29 | 80.7 ± 1.27 | 94.0 ± 0.64 | *Values are percentages. ± denotes run-level SD.* **Feature-BC** replaces directional discrepancy descriptors with ordinary statistics of the merged feature while retaining the same B/C injection. **Constant-BC** removes spatial variation; **Shuffled-BC** preserves map values but disrupts location alignment; **Conflict-Feature** keeps the original discrepancy map while applying direct feature modulation. **B-only** and **C-only** isolate state writing and readout. All tested controls yield lower observed ACC than original joint CCR. A further same-stage intervention reaches **78.9% ACC** (run SD **0.68 pp**), compared with **81.3%** for the proposed next-stage B+C design. Thus, the evidence is not limited to adding an extra branch: it separately tests **conditioning content, spatial correspondence, modulation target, and intervention stage**. These comparisons identify the chosen design as the strongest tested configuration and separately quantify the effects of conditioning content, spatial correspondence, modulation target, and intervention stage.

In the manuscript, the reference to inter-rater disagreement provides intuition for differences between directional responses. The CCR map is presented as a model-derived relational signal rather than a fine-grained clinical lesion annotation, and high-discrepancy locations are not assigned to specific lesion boundaries. Figure 3 provides a qualitative visualization of the model behavior and suggests that CCR may help the model attend to lesion-related regions. The visualization illustrates this potential, while the controlled experiments provide the quantitative evidence for CCR's contribution.

R-w5Gn–Q4. GCRSA center, preprocessing, robustness, and geometric alternatives

Preprocessing and center definition. EUS images were manually cropped to retain the complete esophageal structure. The crop boundaries were determined by organ completeness and did not use probe location; we did not perform probe detection, probe-centered alignment, or anatomical-center estimation. Each cropped image was directly resized to the target resolution without padding. Therefore, the GCRSA reference is the center of the resized image grid and is not guaranteed to coincide with the transducer or an anatomical center. We use it as a deterministic computational reference rather than an anatomical assumption.

GCRSA reorganizes existing feature-grid tokens into circular sequences within the valid inscribed circle and radial sequences across angular sectors. It aggregates clockwise, counter-clockwise, inward, and outward relations through learned relative-offset weights shared across channels, then adds the result through a zero-initialized residual gate. The original Cartesian scans, spatial grid, and token budget remain intact. These are geometric token neighborhoods. The offset weights are input-independent; we do not present GCRSA as input-adaptive attention.

The exact stage-wise grouping and profile configurations are provided in Table G3.

Table R-w5Gn-5. GCRSA structure and alternative implementations on EUS.

ConfigurationACC (%)Recall (%)F1 (%)AUC (%)
VMamba-Small79.9 ± 0.5776.5 ± 2.3079.9 ± 1.0094.1 ± 0.70
Full GCRSA81.1 ± 0.4181.2 ± 0.1081.6 ± 0.8595.0 ± 0.61
Uniform Ring/Ray77.7 ± 1.0976.1 ± 1.5278.1 ± 1.7293.3 ± 0.47
Polar-transformed Input75.6 ± 1.1674.1 ± 2.1176.8 ± 1.9891.1 ± 0.73
Feature-level Polar-SSM73.5 ± 2.2771.1 ± 5.0972.7 ± 3.9290.8 ± 0.76

Values are percentages. ± denotes run-level SD.

The reviewer suggested comparing simpler alternatives that encode related geometric priors, including polar-transformed inputs or features, polar or spiral scan orders inside the SSM, and rotation-equivariant layers. We evaluated the corresponding polar-input, feature-level Polar-SSM, and Uniform Ring/Ray variants. Their ACC values are 75.6%, 73.5%, and 77.7%, respectively, compared with 81.1% for Full GCRSA. By remapping the input or feature coordinates, or by replacing the Cartesian scan order, these alternatives disrupt the spatial arrangement encoded by the pretrained Cartesian weights. GCRSA preserves this pretrained spatial organization and adds circular–radial sensing through a residual branch, strengthening geometric awareness without discarding the original pretrained capability.

We also evaluated the trained GCRSA model under 5% spatial shifts without retraining. The reference displacement and input translation tests evaluate sensitivity to the computational center and image position separately. Moving the GCRSA reference while keeping the input unchanged isolates reference sensitivity; separately translating the input image while holding the reference fixed measures the full model's response to image displacement.

Table R-w5Gn-6. Mean ACC over up/down/left/right 5% perturbations on EUS.

EvaluationACC (%)Change (pp)
Unperturbed GCRSA81.1—
Geometric-reference displacement81.0−0.1
Input image translation80.8−0.3

The geometric-reference displacement reduces mean ACC by only 0.1 pp, and the input-image translation reduces it by 0.3 pp. These small drops indicate that the model retains robust performance under the evaluated geometric perturbations. The image-translation result reflects the response of the full model, including the backbone, while the reference-displacement result directly measures the stability of the GCRSA geometric reference.


Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers

We thank all three reviewers for their careful evaluations and constructive feedback. This response addresses the issues shared across the reviews, including the roles of CCR and GCRSA, improvements over the VMamba-Small backbone on independent medical benchmarks, and the reliability and practical cost of the reported results.

1. Main performance and experimental reliability. Table G1 summarizes the EUS five-fold results, including run-level variability. The Fetal-Planes-DB and OrganAMNIST-224 results are single evaluation protocols and were not obtained through five-fold cross-validation.

Table G1. EUS performance and run variability.

MethodACC (%)Recall (%)F1 (%)AUC (%)
VMamba-Small79.9 ± 0.5776.5 ± 2.3079.9 ± 1.0094.1 ± 0.70
CCR-only81.3 ± 0.5281.5 ± 0.8082.7 ± 0.5593.5 ± 0.57
GCRSA-only81.1 ± 0.4181.2 ± 0.1081.6 ± 0.8595.0 ± 0.61
Full CRGA82.1 ± 0.1680.8 ± 0.5582.9 ± 0.2094.4 ± 0.25

Values are percentages. ± denotes run-level SD.

Full CRGA improves the manuscript-reported accuracy over VMamba-Small by 2.2 percentage points (pp). Both components individually improve the observed EUS accuracy, and their combination gives the highest reported ACC.

Per-run test accuracies and their mean ± sample SD are reported in the detailed statistical response.

2. Specificity of CCR's conditioning information and intervention. CCR measures spatially aligned cross-directional selective-scan response discrepancies before Cross-Merge and uses them to condition subsequent selective-state writing and readout. Its contribution is tested against alternative conditioning sources, spatial controls, and intervention targets under independently trained comparisons.

Table G2. Controlled CCR mechanisms on EUS.

ConfigurationACC (%)Recall (%)F1 (%)AUC (%)
Original CCR (B+C)81.3 ± 0.5281.5 ± 0.8082.7 ± 0.5593.5 ± 0.57
Feature-BC79.0 ± 0.4478.0 ± 3.1479.6 ± 1.8294.0 ± 0.63
Constant-BC80.0 ± 0.6479.3 ± 1.1080.2 ± 1.3593.8 ± 0.69
Shuffled-BC79.3 ± 0.4576.3 ± 1.4579.0 ± 1.0192.8 ± 0.52
Conflict-Feature79.9 ± 0.3176.9 ± 0.5879.3 ± 0.4093.9 ± 0.25
B-only80.1 ± 0.4378.1 ± 1.7280.3 ± 1.3893.3 ± 0.51
C-only80.3 ± 0.9079.4 ± 1.2980.7 ± 1.2794.0 ± 0.64

Values are percentages. ± denotes run-level SD.

Feature-BC replaces cross-directional descriptors with ordinary merged-feature statistics while retaining B/C conditioning. Constant-BC removes spatial variation; Shuffled-BC disrupts spatial correspondence; and Conflict-Feature applies the original discrepancy map directly to features rather than B/C. Original CCR yields the highest observed ACC among these controls. The results identify the combination of directional discrepancy information, spatially aligned conditioning, and state-parameter modulation as the strongest tested configuration. Matched Top-8/Low-8/Random-8 tests further examine the distinct effect of selection strategy at equal spatial budgets.

3. GCRSA's geometric organization and robustness. GCRSA introduces complementary circular and radial relationships between existing tokens while retaining the pretrained Cartesian selective-scan paths. Its grid center is a computational grouping reference, not a presumed transducer, lesion, or anatomical center. On EUS, Full GCRSA obtains 81.1% ACC, compared with 77.7% for Uniform Ring/Ray aggregation, 75.6% for the tested polar-input variant, and 73.5% for the tested feature-level Polar-SSM variant. Under four-direction 5% perturbations, mean ACC is 81.0% for displacement of the geometric reference and 80.8% for image translation, compared with 81.1% without perturbation, showing bounded sensitivity within the evaluated perturbation range.

We appreciate the reviewer’s suggestion. Our original design goal was to preserve the pretrained Cartesian VMamba computation and its initialization mapping. Directly adopting the suggested construction would require changing or remapping this pretrained feature organization, which would conflict with this design constraint.

To make the geometric aggregation fully reproducible, Table G3 reports the exact stage-wise grouping, sequence-length, and four-direction profile configurations used by GCRSA at 224 × 224.

AuthorsOpenReview ↗

General Response to All Reviewers

Table G3. Reference stage-wise GCRSA grouping and profile configuration for 224 × 224 input.

StageFeature gridGCRSA-enabled VSS blocksInscribed-circle radius (grid cells)Number of ringsAngular sectors (configured / non-empty)Ring sequence length (min–max tokens)Radial sequence length (min–max tokens)Profile lengths (CW / CCW / inward / outward)
156 × 5622828176 / 1764–18012–42179 / 179 / 41 / 41
228 × 282141488 / 884–926–2191 / 91 / 20 / 20
314 × 14157744 / 444–443–743 / 43 / 6 / 6
47 × 723.5422 / 201–161–415 / 15 / 3 / 3

The radius, sequence lengths, and profile lengths are measured in feature-grid tokens. “Configured / non-empty angular sectors” denotes the configured number of angular sectors and the number containing at least one assigned token, respectively. The four profile lengths correspond to clockwise, counter-clockwise, inward, and outward propagation, respectively. Profile weights are channel-shared but direction-specific. At Stage 4, 22 angular sectors are configured, of which 20 contain at least one assigned token.

Geometric indices and valid-position masks are precomputed for each stage. Circular and radial aggregation use parallel masked accumulation and normalization with no polar resampling or additional tokens. The values in Table G3 are the reference instantiation for 224 × 224 input, rather than fixed resolution-specific constraints. For other input resolutions, the feature-grid sizes, ring and sector counts, valid-position masks, sequence lengths, and profile lengths are regenerated deterministically from the same grouping rule. Shape-adapted offset profiles are used during evaluation without additional training.

To address the reviewer’s concern while retaining the pretrained Cartesian paths, we conducted additional controlled experiments under the same training and evaluation protocol. These experiments show clear performance degradation relative to the proposed residual circular–radial design. We will provide the exact experimental configurations and quantitative results in the revised manuscript.

4. Public benchmark evidence. We have added the unmodified VMamba-Small backbone on both public datasets. Full CRGA improves ACC from 94.8% to 95.5% (+0.7 pp) on Fetal-Planes-DB and from 95.2% to 96.3% (+1.1 pp) on OrganAMNIST-224. These same-backbone results establish the observed improvement of the complete CRGA architecture outside the primary EUS task. Additional CCR-only and GCRSA-only results address component attribution in the corresponding reviewer responses; the public benchmarks principally assess the full model's competitiveness and broader medical-task applicability.

5. Statistical interpretation and measured computational cost. We additionally performed patient-level paired comparisons with exact McNemar tests on the shared EUS test cohort and measured inference latency, peak GPU allocated memory, and resolution scaling. The former are reported separately from run-level variability; the latter reveal a material runtime overhead not captured by FLOPs alone. At 224 × 224, batch size 1, measured model-forward latency is 9.1 ms for VMamba-Small versus 13.7 ms for Full CRGA, with peak allocated memory 223.0 MB versus 225.1 MB. The reviewer-specific responses give the paired uncertainty estimates, complete resource tables, and implementation details.


From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.