acceptodds
Under review as a conference paper at ICLR 2027

Establish, Break, Rebuild, Consolidate: Relational Self-Supervised Learning for Cross-Dataset Skin-Lesion Representation Learning

Abstract

Skin-lesion representations should preserve clinically meaningful relationships across acquisition conditions, yet dermoscopic and clinical or smartphone images can differ substantially in illumination, scale, viewpoint, color response, and visible structure. We study self-supervised representation learning from independently collected and unpaired datasets without using diagnosis labels to define cross-dataset correspondences. We introduce Cross-dataset INter-view relational Distillation with Evidence-masked Recovery and Stabilization (CINDERS), which learns representations by explicitly modeling relationships among samples. An exponential-moving-average teacher defines the relational organization of intact multi-view images and ranks local visual evidence. A shared student first establishes this intact relation; the highest-ranked patches are then physically removed before a second frozen backbone pass; the student rebuilds the intact relation from the surviving evidence; and a symmetric contrastive objective consolidates the intact and rebuilt representations. We evaluate the learned representation primarily through cross-acquisition retrieval at rank 1 (R@1). On DermaMNIST-C and PAD-UFES-20, CINDERS improves the two cross-acquisition R@1 directions from for frozen DINOv3 to , increasing their harmonic mean from to . Under the matched self-supervised learning protocol, CINDERS achieves a four-direction mean R@1 of , compared with for iBOT, while its cross-acquisition harmonic mean of exceeds the obtained by Relational Knowledge Distillation. On the external Derm7pt–MILK10K evaluation, the four-direction mean R@1 increases from for frozen DINOv3 to , and the cross-acquisition harmonic mean increases from to . Together, the internal and external results show that recovering inter-sample relationships from incomplete visual evidence can improve cross-acquisition representation compatibility without diagnosis supervision or manually specified cross-dataset positive pairs.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
self-supervised learning, representation learning, relational learning, cross-dataset retrieval, domain shift, medical imaging, dermatology, skin lesion analysis

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_WWRxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper's proposition is the skin lesion representations should preserve clinical meaningful relationships across different acquisitions method and condition. With self-supervised representation learning as the motivation, they introduced Cross-dataset INter-view relational Distillation with Evidence-masked Recovery and Stabilization (CINDERS), with a teacher-student model. The main contribution is formulating self-supervised learning as relational recovery from incomplete evidence, demonstrated through a shared student, and evaluated learned representation through internal and external cross-acquisiation retrieval.

Critical strengths

The proposed Establish-Break-Rebuild-Consolidate in the method, an interesting idea in representation learning. Formulating SSL as relational recovery from incomplete evidence, demonstrated through a shared student, and evaluated learned representation through internal and external cross-acquisiation retrieval.

Critical weaknesses

The research questions can be clearer and improve the clarity of the association with the results reported. The review and references are not up-to-date, particularly, Table 1 is comparing with dated method 2020-2022, how about recent works between 2023-2026. It is promising to demonstrate the learned representation through internal and external cross-acquisition retrieval, but lack evidence of this improves the recognition rate. The evaluation lack robustness and rigour.

Questions for authors

Table 1 is comparing with dated method 2020-2022, how about recent works from CVPR 2026? With the improved of learned representation (or your results section), how do these perform on cross-domain in skin lesion recognition? Can this method be generalized to other domain, like natural objects?

3 comments

Thank you for your constructive comments. We clarify the research questions and provide experiments on training objectives, masking, staging, and external recognition.

Weakness 1 (research questions). We study three questions: (1) Does recovering the intact teacher’s relational structure from incomplete visual evidence improve retrieval across both within-acquisition (D→D, C→C) and cross-acquisition (D→C, C→D) settings? (2) How do the strategy and extent of evidence removal affect relational representation learning? (3) Does temporal staging improve optimization beyond relational Rebuild?

To evaluate the learning objective, we compared relational Rebuild, MAE, and feature matching under two mask-selection strategies: the 40 highest-attention patches versus 40 randomly selected patches. Each setting used 10 runs of 50 training epochs. Relation achieved Mean4/Cross-HM of 58.05/44.25% (attention) and 58.40/44.46% (random), versus MAE’s 52.67/34.27% and 53.51/36.02%. The Cross-HM improvements over MAE were +9.98/+8.45 points, with p=1.31×10−11/1.78×10−10p = 1.31 \times 10^{- 11}/1.78 \times 10^{- 10}, respectively. A separate 10-run staging comparison gives a +1.41-point Cross-HM improvement (p=0.0306p = 0.0306).

Weakness 2 (recent methods, 2023–2026). We expand the related-work discussion to Relational Representation Learning (Maes et al., 2025), X-Sample Contrastive Loss (Sobal et al., ICLR 2025), and Relational Visual Similarity (CVPR 2026). These provide important precedents for relational learning; the latter studies relations among elements within images, whereas CINDERS learns to recover relationships between images after evidence removal across unpaired acquisition domains. Our SSL comparisons train the DFF using each method’s objective while the pretrained DINOv3 backbone remains frozen.

To distinguish our relational objective from conventional distillation, we compared it with similarity-KL and feature matching using attention-selected masking of 123 patches, same-state optimization, and 10 runs of 100 training epochs:

RebuildMean4 R@1Cross-HM R@1
Relation64.22±1.2853.54±2.75
Similarity KL57.79±0.8542.52±1.41
Feature matching56.45±0.7040.55±1.42

Relation improves Mean4 by 6.43/7.77 points and Cross-HM by 11.02/12.99 points over similarity-KL/feature matching.

Weakness 3 (skin-lesion recognition). We evaluated linear-probe classification on external MILK10K and DDI datasets. We froze the representation network, trained a linear classifier on labeled training samples, and evaluated held-out diagnostic AUC using the 10 runs with random selection of 40 masked patches:

Rebuild objectiveMILK10K AUC (%)DDI AUC (%)
Relation71.56±0.7861.62±2.77
MAE71.38±0.6661.39±1.61
Feature matching70.69±0.7860.47±2.73

Relational Rebuild achieves the highest mean AUC in these two datasets. Differences are modest, so retrieval remains the principal evaluation outcome. Unlike classification, which only requires representations to separate diagnostic classes, retrieval requires preserving fine-grained inter-image relationships while bridging substantial appearance differences across acquisition modalities, making it a more demanding test of the relational representation targeted by our method.

Weakness 4 (robustness and masking policy). We compared teacher-attention selection against random masking for 40, 96, and 123 patches out of 196, with each experiment trained for 100 epochs. The table reports fine-diagnosis retrieval R@1 (%), mean ± standard deviation across runs:

MaskingRunsD→C R@1C→D R@1Mean4 R@1Cross-HM R@1
Top40649.39±2.6837.88±5.1757.82±1.5342.75±3.84
Random40639.56±1.4752.68±1.3859.28±0.7145.17±1.19
Top96363.84±4.1652.61±1.4565.27±1.0657.67±2.49
Random96344.96±2.2956.22±2.5161.46±0.9549.95±2.17
Top123662.04±4.6845.78±2.0363.24±1.6152.66±2.96
Random123652.73±2.7553.55±1.8363.32±0.9653.11±1.91

With 40 masked patches, Top improves D→C by 9.83 points but reduces C→D by 14.79 relative to Random; Random has higher Mean4 and Cross-HM. With 96 patches, attention selection improves both aggregate metrics; with 123, Top and Random are comparable overall. This inconsistent evidence across masking ratios means we do not claim that attention-based mask selection is universally favorable.

Question (natural-image generalization). Relational Representation Learning demonstrates relational modeling on CIFAR-10, CIFAR-100, and ImageNet-100, supporting the broader principle beyond medical images. CINDERS requires no lesion diagnosis labels or paired lesions for training. However, that prior study does not test CINDERS’s Break–Rebuild mechanism; experiments on natural objects remain an important future direction.

Reviewer_WWRxOpenReview ↗

Result for natural image generalisation

Thanks for your response. Could you please outline what are the changes on the revised manuscript (if you have one, please upload). The generalisation aspect is not clear and no results reported in the authors' responses. I'll keep my score for now.

Thank you for your follow-up. We have prepared a revised manuscript and supplementary material, with the modifications highlighted in blue. We summarize the changes below.

1. Changes in the revised manuscript

  • Introduction (p. 2): Refined the contributions to distinguish relational Rebuild, evidence removal, and temporal staging.
  • Related Work (p. 3): Expanded the discussion of recent relational self-supervised learning and added cross-modal representation learning.
  • Method (pp. 4–7): Clarified the relational objectives, teacher–student input formations, physical patch-token removal, and temporally staged optimization.
  • Controlled ablations (Table 2, p. 8): Added multi-seed experiments comparing relational Rebuild, MAE reconstruction, and feature matching under attention-guided and random masking. These comparisons use ten runs per setting and directly examine whether relational recovery provides improvements beyond conventional reconstruction and feature distillation.
  • External recognition (Table 5, p. 10): Expanded the linear-probe evaluation to report precision and recall alongside AUC on MILK10K and DDI.
  • Supplementary material: Added detailed experiments on relational versus similarity-KL distillation, within- versus cross-acquisition relationships, masking ratios, temporal staging, and class-wise recognition performance.

2. Clarification of generalization evidence

The external Derm7pt–MILK10K retrieval evaluation was already included in the original manuscript (Table 4). We retain this experiment in the revised version and highlight its results here to clarify the generalization evidence.

Unlike the internal DermaMNIST-C/PAD-UFES-20 evaluation, this experiment measures retrieval between two independently collected external datasets. Each query searches the other dataset, with the evaluation covering both within-acquisition and cross-acquisition retrieval.

Compared with frozen DINOv3, CINDERS achieves:

  • Dermoscopic-to-clinical R@1: 23.63% to 36.72%.
  • Clinical-to-dermoscopic R@1: 29.87% to 35.72%.
  • Four-direction mean R@1: 29.61% to 36.65%.
  • Cross-acquisition harmonic mean: 26.38% to 36.21%.

These existing results demonstrate improved representation compatibility across external dermatology datasets and acquisition conditions. They complement the newly added controlled ablations, which isolate the contribution of relational Rebuild.

3. Scope of generalization

Our experimental evidence supports generalization across independently collected skin-lesion datasets and clinical/dermoscopic acquisition conditions. It does not establish generalization to unrelated natural-image datasets, which remains an important direction for future evaluation.

We hope this summary clarifies the manuscript revisions and the distinction between the newly added experimental evidence and the external generalization results already reported in the original submission.

Reviewer_HmLnOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces CINDERS for learning skin-lesion representations across dermoscopic and clinical images without diagnosis-defined training pairs. The method combines teacher-guided patch removal, relational recovery and contrastive alignment through a staggered training schedule. Results show better bidirectional retrieval than frozen DINOv3 and several matched objective controls.

Critical strengths
  1. The proposed method does not need paired dermoscopic and clinical images. This makes the setting practically relevant, since independently collected datasets may not have matching lesions.

  2. Removing patches before the second backbone pass is a sensible design choice, ensuring the remaining tokens cannot access the removed content through attention during that pass.

  3. Testing on external datasets provide some evidence that the result is not specific to the internal datasets.

Critical weaknesses
  1. In Table 2, MAE-only Rebuild actually achieves a higher Mean4 than the full method, 0.6068 vs. 0.6014. Its harmonic mean is also close, 0.5046 vs. 0.5128. The full method gives better directional balance, but I’m not convinced this establishes the need for relational Rebuild. A simpler reconstruction-based approach appears competitive, and this comparison deserves more attention. Thus, the key contribution is not sufficiently supported by the ablation results.

  2. Using the same backbone, fusion module and training horizon is useful, but it does not necessarily mean the methods receive the same compute or tuning budget. CINDERS has repeated backbone processing and several objectives. The same-state ablation is relevant, but there are still some detail missing about whether it isolates scheduling while keeping the other factors comparable.

  3. The paper does not clearly establish what is new beyond combining existing ideas. Teacher–student learning, relational distillation, masking and contrastive alignment all have substantial prior work. The authors acknowledge this, but the related-work discussion does not go far enough in explaining why their particular combination constitutes a meaningful methodological advance. A particularly relevant omission is AttMask (ECCV 2022), which already uses teacher attention to select what to hide from the student. ADIOS (ICML 2022) also studies challenging masks together with consistency between original and masked representations.

Questions for authors
  1. How does Break differ from AttMask’s teacher-attention-guided masking? Is the key distinction physical token removal, the relational target, or something else? Which experiment isolates that difference?

  2. What does recovering inter-sample relations add beyond matching teacher features from incomplete inputs, as explored in related masked-prediction and distillation methods?

  3. Could attention-guided masking plus a standard relational-distillation objective achieve similar results without the four-stage schedule? Can the authors test this under comparable compute and tuning budgets?

  4. What is the main novelty claim: the relational recovery objective, the temporal schedule, or their integration? The paper would benefit from stating this more precisely.

  5. Given the competitive MAE-only result, what evidence shows that relational recovery is essential to the improvement rather than one of several workable objectives?

3 comments

Thank you for the review. We separate Rebuild and scheduling effects.

Weakness 1 (MAE-only versus relational Rebuild). The original MAE-only run leads Mean4 R@1 (60.68% vs 60.14%); CINDERS leads Cross-HM (51.28% vs 50.46%). We separate mask selection, Rebuild objective, and run variation: relation/MAE/feature with 40 attention-selected or random masked patches, 10 runs of 50 epochs each.

MaskingRebuildD→C R@1C→D R@1Mean4 R@1Cross-HM R@1
Top40Relation49.27±3.3140.24±1.4358.05±0.9644.25±1.61
Top40MAE32.55±2.0536.31±1.5352.67±0.3634.27±1.10
Top40Feature32.85±1.8937.07±2.1752.85±0.7134.79±1.58
Random40Relation40.99±1.7348.67±1.6458.40±0.6944.46±1.08
Random40MAE35.96±1.2736.18±2.8253.51±0.7836.02±1.61
Random40Feature34.58±1.5337.15±2.1353.32±0.6735.76±1.03

Relation improves Mean4 by 4.89/5.38 points and Cross-HM by 8.45/9.98 points over MAE/feature. For Cross-HM, Top40 relation versus MAE/feature gives p=1.31×10−11/5.04×10−11p = 1.31 \times 10^{- 11}/5.04 \times 10^{- 11}; Random40 gives p=1.78×10−10/2.02×10−13p = 1.78 \times 10^{- 10}/2.02 \times 10^{- 13}. The gains persist across runs.

Weakness 2 (objective and scheduling). We conducted two complementary analyses.

First, with same-state optimization, we compared relation, similarity-KL and feature distillation, masking 123/196 attention-selected patches over 10 runs of 100 epochs:

Recovery objectiveD→C R@1C→D R@1Mean4 R@1Cross-HM R@1
Proposed relation67.05±4.5444.94±4.9064.22±1.2853.54±2.75
Similarity KL45.03±2.0040.36±2.0857.79±0.8542.52±1.41
Feature distillation42.21±2.2739.08±1.4456.45±0.7040.55±1.42

Without staging, relation improves Mean4 by 6.43/7.77 and Cross-HM by 11.02/12.99 points over similarity-KL/feature (p=1.58×10−8/1.96×10−9p = 1.58 \times 10^{- 8}/1.96 \times 10^{- 9} for Cross-HM).

Second, we compared staged versus no-staging in 10 runs of 50 epochs:

TrainingD→C R@1C→D R@1Mean4 R@1Cross-HM R@1
Staged47.73±4.7341.45±1.9757.97±1.1844.22±2.39
No-staging46.22±3.4140.08±2.9957.34±1.1942.81±2.25

Staging improves Cross-HM by 1.41 points (p=0.0306p = 0.0306) across ten runs. Relational Rebuild delivers the larger gain; staging adds benefit.

Weakness 3 (novelty). We acknowledge that existing teacher–student learning, AttMask, ADIOS, and relational distillation provide important foundations. Our novelty lies in relational rebuild for cross-acquisition retrieval rather than reconstructing masked content, Break deliberately removes visual evidence and Rebuild trains the student to recover the intact teacher’s within- and cross-acquisition relationships. Consolidate and staged training further stabilize this relational objective.

Q1 (AttMask). AttMask uses teacher attention to select masking regions. CINDERS removes selected tokens before the second frozen-DINO pass. Masking breaks relationships; relational Rebuild is the learning objective. The masking/objective experiment verifies its gains under both masking strategies.

Q2 (versus feature matching). Feature matching supervises embeddings; similarity-KL distills unmodified similarity distributions. Our objective removes acquisition-block mean offsets, distills normalized inter-image relations, and preserves cosine geometry. The same-state results verify the distinction.

Q3 (standard similarity distillation without staging). In the same-state 123-patch setting, similarity-KL yields 57.79% Mean4/42.52% Cross-HM versus 64.22%/53.54% for relational Rebuild. The second table tests staging.

Q4 (novelty). The central contribution is relational Rebuild as a learning objective for cross-acquisition retrieval. An intact EMA teacher establishes relationships among dermoscopic and clinical images without labels or paired lesions. Break removes image evidence to perturb those relationships, and Rebuild trains the student to recover the teacher’s relational structure. Consolidate uses both clean and masked representations, while temporal staging coordinates learning over successive model updates. The novelty lies in this relational learning objective and its role in cross-acquisition representation learning, not in attention-based masking alone.

Q5 (importance). In the 10-run, 50-epoch, 40-patch masking study (Weakness 1), relational Rebuild achieves Cross-HM 44.25% versus MAE 34.27% and feature matching 34.79% with attention masking; with random masking, the corresponding scores are 44.46% versus 36.02%/35.76%. In the 10-run, 100-epoch, 123-patch same-state study (Weakness 2), it reaches 53.54% versus similarity-KL 42.52% and feature distillation 40.55%. These settings specifically establish why relational Rebuild is our principal learning objective for cross-acquisition retrieval.

Thank you for the detailed response. The additional experiment helps to explain the technical contributions. Justification about novelty is helpful in how it builds on existing T-S concept. I have revised my score, but some concerns remain unresolved.

Thank you for reconsidering your assessment and for recognizing the additional experimental evidence and clarification of our technical contributions.

We have also prepared a revised manuscript and supplementary material, with all modifications highlighted in blue, to incorporate these clarifications directly into the paper.

In particular, we have reframed the contributions in the Introduction (p. 2) to distinguish relational Rebuild, evidence removal, and temporal staging, while explicitly acknowledging existing teacher–student and masking approaches in Related Work (p. 3).

More importantly, Table 2 (p. 8) now includes the new controlled, multi-seed ablations comparing relational Rebuild against MAE reconstruction and feature matching under both attention-guided and random masking. The supplement further includes comparisons against conventional similarity-KL distillation without temporal staging, as well as separate analyses of masking strategies and staging effects.

These revisions aim to clarify that our central contribution is recovering intact inter-image relational structure from incomplete visual evidence, rather than introducing teacher–student distillation or attention-based masking in isolation.

We appreciate that some concerns remain and would welcome any further clarification on the specific points that you feel are not yet sufficiently addressed. Thank you again for your constructive feedback.

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

The paper presents CINDERS, a self-supervised learning method for cross-dataset skin-lesion representation learning based on four components. Establish establishes relations among the original (intact) skin-lesion images. Break removes informative regions of the images selected based on attention while Rebuild restores the inter-sample relational distribution from evidence-reduced views. The last component, Consolidate aligns then the intact and rebuilt representations. Learning uses temporally staggered optimization. Experiments are performed on DermaMNIST-C and PAD-UFES-20. A cross-dataset evaluation on Derm7pt and MILK10K is also performed.

Critical strengths
  • The proposed method makes use of a reasonable combination of machine learning techniques to recover a representation that bridges the dermoscopic-clinical imaging gap in both directions.
  • It attempts to do this in an unsupervised manner, i.e., without using diagnosis labels.
Critical weaknesses
  1. The clarity and presentation quality of the paper can be largely improved. The figures are of low quality (low resolution rasters) and introduce ambiguities. Regarding clarity, not all symbols are defined. Examples are provided in the detailed comments bellow.

  2. The motivation behind the proposed method stems from cross-acquisition (imaging modality) retrieval but Section 4.4 is quite brief and does not discuss in sufficient detail how the proposed method achieve this goal. A deeper theoretical or empirical analysis should be included, supporting how relational distillation across a balanced batch alone is sufficient to bridge the domain shift.

  3. The results (Table 1) show that the proposed method excels in balanced bi-directional performance on the metrics considered by the authors, but has reduced performance on Mean R@1R@1 and RDC@1R_{DC}@1.

  4. Related to the previous point, the motivation behind bridging both gaps, i.e., dermoscopic to clinical AND clinical to dermoscopic needs to be better supported. The authors should describe practical scenarios and applications that highlight the importance of both cross-acquisition directions.

  5. Cohorts are not properly defined.

Detailed comments

  • L.024: "physically removed" doesn't sound right.
  • Figure 2: "... provides a detached teacher feature together with patch attention ..." needs clarification.
  • L.212: The definition of intact images is provided here but the term is used already from the abstract.
  • CECE in LNCEL_{NCE} is not defined
  • Figure 2 does not depict what is stated at ziEz_i^E definition in eq. 1 as augmented versions of the images provided to Establish are shown (though the captions state "Original (to {o,g,c,l})".
  • L.313-314: Please elaborate on this statement (controls over reproductions).
  • Table 1: What does "Random DFF initialization" method corresponds to?
  • L.413-419: The evaluation protocol followed in this paragraph is not clear. Additionally, the corresponding notation is confusing, mixing XD standing for cross-dataset with D as "dermoscopic" and {X,Y} acting as placeholders for {C,D}.
Questions for authors

Please address the points raised in the weaknesses above.

4 comments

Thank you for the review. We address the concerns.

Weakness 1 (presentation). We improve figures and notation.

Weakness 2 (cross-acquisition mechanism). Our objective preserves relationships within and across acquisitions. Each batch has 64 dermoscopic and 64 clinical images, yielding 4,032 within-acquisition and 4,096 cross-acquisition image pairs. Balanced sampling alone does not align domains. Our EMA teacher constructs inter-image similarities, removes acquisition-dependent mean offsets, and guides the student to recover relations from incomplete evidence.

We compare within-acquisition (D→D, C→C), cross-acquisition (D→C, C→D), and combined (D→D, C→C, D→C, C→D) relations. With staged training for 100 epochs masking the 123 highest-attention patches of 196, within-only achieves 62.57% Mean4 R@1 and 51.13% Cross-HM versus cross-only (59.01% and 46.24%). Combined relations perform best (63.24% Mean4; 52.66% Cross-HM).

Relational objectiveRunsMean4 R@1Cross-HM R@1
Within-acquisition only1062.57±2.13%51.13±4.21%
Cross-acquisition only1059.01±2.39%46.24±4.95%
Full (within + cross)1063.24±1.61%52.66±2.96%

These findings show that within-acquisition geometry matters for cross-acquisition retrieval; combining both relations further improves learning without diagnosis labels or paired images.

Weakness 3 (metric trade-offs). Versus iBOT, CINDERS has lower D→C (53.30% vs 77.04%) but higher C→D (49.40% vs 13.65%), Mean4 (60.14% vs 55.27%) and Cross-HM (51.28% vs 23.19%). The original MAE-only run leads Mean4 (60.68% vs 60.14%). We compare Rebuild objectives with 40 attention-selected masked patches in 10 runs of 50 epochs per setting:

Rebuild objectiveD→CC→DMean4Cross-HM
Relation49.27±3.3140.24±1.4358.05±0.9644.25±1.61
MAE32.55±2.0536.31±1.5352.67±0.3634.27±1.10
Feature matching32.85±1.8937.07±2.1752.85±0.7134.79±1.58

Relation improves Mean4 by 5.38/5.20 points and Cross-HM by 9.98/9.45 over MAE/feature (p=1.31×10−11p = 1.31 \times 10^{- 11} versus MAE; p=5.04×10−11p = 5.04 \times 10^{- 11} versus feature matching). This supports relational Rebuild without dismissing the original MAE result.

Weakness 4 (both directions). The two directions are complementary: clinical and dermoscopic images capture different manifestations of the same lesion. Clinical images show macroscopic appearance and context, whereas dermoscopy reveals subsurface diagnostic structures. Clinical→dermoscopic tests whether macroscopic features retrieve relevant dermoscopic phenotypes; dermoscopic→clinical tests whether those phenotypes generalize to heterogeneous real-world appearances. Evaluating both tests diagnosis-level correspondence rather than modality-specific similarity.

Weakness 5 (cohorts). A cohort is a balanced dermoscopic/clinical batch tracked across stages. At iteration tt, Establish constructs relations for cohort tt, Rebuild learns from masked cohort t−1t - 1, and Consolidate revisits clean and masked views of cohort t−2t - 2. Their losses combine into one student update, followed by an EMA-teacher update. Each cohort experiences successive stages and model states, not all stages simultaneously.

Detailed comments. L.024: Teacher attention selects the top 40 of 196 patches. Their tokens are removed before the second frozen-DINO pass; the remaining 156 retain their positions, without replacements. Figure 2: The EMA teacher supplies stop-gradient relation targets and separate masking-attention scores. L.212: Intact means unmasked, not unaugmented.

Notation. ii: sample index; vv: view (oo: original, gg: geometry, cc: color, ll: lighting); pp: patch index (1--196); xivx_i^v: image view; HivH_i^v: frozen-DINO feature; AT,iv(p)A_{T,i}^v(p): teacher attention; Iiv\mathcal{I}_i^v: removed-patch indices; ziE,ziRz_i^E,z_i^R: Establish/Rebuild embeddings; T,ST,S: teacher/student; sgsg: stop-gradient; τ\tau: temperature; LNCEL_{NCE}: InfoNCE loss.

Figure 2/Eq. 1: The teacher uses four intact views; Establish repeats the original representation in four DFF slots; Rebuild uses four masked views. L.313–314: Compared SSL methods train the DFF with their objectives; pretrained DINOv3 stays frozen. Table 1: DFF uses warmup distillation from frozen DINOv3; Random DFF skips it. L.413–419: Let X,Y∈{D,C}X,Y \in \{ D,C\} denote query/gallery acquisition, 77 Derm7pt and MM MILK10K: RXYXD@1=12[R@1(X7→YM)+R@1(XM→Y7)]R_{XY}^{XD}@1 = \frac{1}{2}\left\lbrack R@1\left( X_{7} \rightarrow Y_{M} \right) + R@1\left( X_{M} \rightarrow Y_{7} \right) \right\rbrack. DD, CC, DC, CD average both dataset-origin directions. Each query searches the opposite dataset and requires a diagnosis match with a positive in the gallery.

Thank you again for your constructive review. We would like to highlight that we have prepared a revised manuscript and supplementary material, with the modifications highlighted in blue, incorporating the clarifications and additional experiments discussed in our previous response.

In particular, we have made the following revisions to address your concerns:

  1. Presentation and methodology (Figures 1–2, pp. 4–7): Clarified the notation, teacher–student architecture, physical patch-token removal, and the Establish–Break–Rebuild–Consolidate training sequence. We also explain how within- and cross-acquisition relationships are constructed from balanced batches without diagnosis labels or paired lesions.

  2. Cross-acquisition mechanism (Section 5.3 and Supplement): Added controlled experiments separating within-acquisition and cross-acquisition relational objectives, alongside the combined configuration. These experiments clarify how the different relational components contribute to bidirectional retrieval.

  3. Retrieval trade-offs and controlled ablations (Table 2, p. 8): Expanded the component analysis with ten-run comparisons of relational Rebuild, MAE reconstruction, and feature matching under both attention-guided and random masking. Relational Rebuild consistently improves the aggregate retrieval metrics in these matched experiments, addressing the concern that individual checkpoint results may favor MAE or show directional trade-offs.

  4. Bidirectional evaluation (Section 5): Clarified the distinct roles of dermoscopic-to-clinical and clinical-to-dermoscopic retrieval, and why both directions are necessary to assess cross-acquisition representation compatibility rather than relying on a single retrieval direction.

  5. Additional experimental evidence (Supplement): Included detailed analyses of relational objectives, masking strategies, temporal staging, and their effects on retrieval performance.

We hope these revisions make the technical mechanism, experimental controls, and interpretation of the results clearer. In particular, the new controlled ablations strengthen the evidence that the improvements arise primarily from relational recovery rather than masking or teacher–student distillation alone.

We sincerely appreciate your feedback and hope the revised manuscript addresses your remaining concerns.

I thank the authors for adding the multi-seed experiments and clarifying the acquisition-block centering. However, the revised results raise two concerns:

  1. Why was the within vs. cross-acquisition ablation (L.462–465) evaluated using a different protocol (100 epochs and 123 masked patches compared to 50-epoch, 40-patch setup)? It seams that across Sections 5.3 and Section 6, different claims rely on different experimental settings making comparison difficult.
  2. Under the standard 50-epoch, 40-patch protocol in Table 2B, the reported numbers differ significantly from those reported in Table 2A. Could the authors clarify the source of this discrepancy?

Finally, several of my previous minor comments remain unaddressed in the revised manuscript.

Thank you for the follow-up questions. We appreciate the opportunity to clarify the experimental design and the differences between the reported results.

1. Different training protocols for the within- and cross-acquisition ablation

The experiments address two different questions.

Table 2B is our primary controlled comparison of Rebuild objectives. We compare relational Rebuild, MAE reconstruction, and feature matching under attention-guided and random masking, using 50 epochs, 40 masked patches, and 10 runs per setting. This experiment isolates the contribution of relational recovery under a consistent protocol.

The within- versus cross-acquisition ablation examines a different aspect: whether learning relationships within the same acquisition domain or across acquisition domains contributes more to the representation. We use 100 epochs and 123 masked patches, with 10 runs for each relational objective. The training settings are identical within this comparison, allowing us to isolate the effect of the relational components.

These are two separate controlled studies rather than results intended for direct numerical comparison. We will clarify their respective purposes and settings in Sections 5.3 and 6, distinguishing conclusions about the Rebuild objective from those about relational components.

2. Differences between Tables 2A and 2B

Table 2A presents the original component ablations using selected model checkpoints, whereas Table 2B reports the newly added controlled experiments with 10 runs of 50 epochs under a consistent training and evaluation protocol.

For example, the full CINDERS configuration achieves Mean4/Cross-HM of 60.14/51.28% in Table 2A, compared with 58.38 ± 1.10/45.04 ± 2.06% in Table 2B. The panels use different experimental and checkpoint-selection procedures, so their absolute values are not directly interchangeable.

The purpose of Table 2A is to examine the contributions of individual framework components, while Table 2B provides a more systematic comparison of Rebuild objectives. In Table 2B, relational Rebuild improves Cross-HM over MAE by 9.98 points with attention masking and 8.45 points with random masking, consistently supporting the contribution of relational recovery.

We will make the distinction between the two tables more explicit in the manuscript, including their experimental settings and evaluation procedures, to prevent confusion.

Thank you again for these suggestions. We believe these clarifications will make the experimental comparisons and their conclusions easier to follow.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.