{"id":"8b00970f-d0f5-44a6-a04d-0bbf2c13f0c6","arxiv_id":"2509.10134","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Grad-CL uses Grad-CAM features for pseudo-label refinement and a cosine contrastive loss to disalign optic cup and disc features, reporting improved segmentation results.","lead":"Grad-CL is a source-free domain adaptation method for optic disc and cup segmentation that refines noisy pseudo-labels with gradient-guided features and adds a cosine similarity contrastive loss to separate class-specific features. It reports higher Dice scores than prior state-of-the-art SFDA methods on RIM-ONE-r3 and Drishti-GS with REFUGE as source.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Grad-CL's mechanism rests on source-model Grad-CAM maps for target images, yet the paper never checks that these maps actually localize cup and disc under domain shift; if they don't, both pseudolabel refinement and the contrastive loss amplify the wrong signal.","rationale":"The reader's weakest assumption is exactly that the gradient maps reliably locate optic cup and disc on target images. I agree this is the load-bearing point: every component that distinguishes Grad-CL from DPL-style denoised pseudolabeling—gradient-guided prototype refinement and gradient-derived contrastive disalignment—is built on these maps. If the maps are unreliable under domain shift, the method does not implement what the paper claims, and the reported improvements over DPL cannot be attributed to the proposed mechanism. The paper gives no evidence on this point, and its Table 2 'ablation' tests distance metrics rather than localization quality, so it does not speak to the assumption. My proposed check is feasible because target ground-truth masks are already used for evaluation; computing map Dice against them is a diagnostic, not a violation of source-free adaptation. Since this is a serious but empirically testable concern, the correct stance is conditional acceptance pending the localization check, which matches the existing verdict. I do not see a reason to move the verdict.","tokens_in":11024,"tokens_out":8362,"duration_ms":105249,"concrete_test":"Using the same REFUGE/Drishti/RIM-ONE splits as in Section 4, take the pretrained source model and compute e_cup_GC and e_disc_GC for all target validation images. Threshold each map (e.g., Otsu or top-k percent) and compute Dice/IoU against the target ground-truth cup and disc masks; do the same for source validation images. If target map Dice is far below source map Dice or near chance, the refinement and contrastive targets are unreliable and the central mechanism is unsupported. If target map Dice is comparable to source, the gradient-localization assumption is validated and the main concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on Eqs. (4)-(5): e_cup_GC and e_disc_GC are computed from gradients of the pretrained source model on target images. These maps directly enter prototype estimation (Eqs. 6-8), the denoising mask (Eq. 9), and the contrastive loss (Eqs. 11-12). For the reported improvement to be caused by the proposed mechanism, these maps must highlight the optic cup and disc reliably on target data and remain class-discriminative under domain shift. That is precisely the condition domain shift threatens. The paper provides no target-domain validation: no Grad-CAM visualizations on target images, no quantitative localization test against target ground-truth masks, and no comparison with source-domain map quality. If the maps are corrupted, Eq. 9 will discard correct pseudolabels and Eq. 12 will push the wrong features apart, so the gains could come from the extra training signal rather than from the proposed gradient-guided refinement. This is not an internal inconsistency, but it is the unverified load-bearing assumption. Table 2 does not resolve it: it compares distance/divergence metrics on the maps, not their localization accuracy.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Grad-CL, a source-free domain adaptation method for optic cup and disc segmentation in fundus images. It consists of two stages: (i) gradient-guided pseudolabel refinement, where Grad-CAM-like maps computed from the pretrained source model on target images are used to modulate penultimate-layer features, build prototypes, and construct a refinement mask for pseudolabels; and (ii) contrastive feature disalignment, where cosine similarity between gradient-informed cup and disc features is minimized. The method is evaluated on REFUGE→RIM-ONE-r3 and REFUGE→Drishti-GS, reporting improvements over several unsupervised and source-free baselines. The paper also includes an 'ablation' table comparing divergence versus distance metrics on gradient maps.","tokens_in":11396,"tokens_out":6641,"duration_ms":75936,"significance":"If the mechanism works as claimed, Grad-CL would be a useful contribution to source-free segmentation in medical imaging, particularly for tasks with small, structured anatomical regions. The paper releases code and evaluates on two public benchmarks against multiple UDA and SFDA baselines, which is a strength. However, the core contribution is currently underspecified: the central operation involving gradient-modified features is dimensionally ambiguous, the contrastive loss is not fully defined, the loss weight λ is missing, and the 'ablation' does not test the proposed components. The reported gains are also small in several comparisons and are presented without error bars. The most serious concern is that the entire method rests on the assumption that source-model Grad-CAM maps localize the cup and disc on target images, an assumption that is never validated on target data and could be violated by domain shift.","major_comments":[{"comment":"The quantity e_GCv is never defined. Eq. (5) defines e_cup_GC and e_disc_GC as spatial heatmaps, but Eq. (6) multiplies this quantity by the penultimate-layer feature vector e_v, and Eq. (11) adds it to e_v. If e_GCv is a scalar heatmap value, the vector addition is dimensionally inconsistent; if it is a per-pixel feature vector, its construction is missing. Since every component depends on this operation, the definition and dimensionality must be stated explicitly and the equations must be made consistent.","section":"Section 3.1, Eqs. (5)-(8); Section 3.2, Eq. (11)"},{"comment":"The refinement mask only distinguishes object vs. background, although the segmentation task has two foreground classes (cup and disc). The relationship between the binary object indicator \\hat{y}=1 and the two logits y_cup, y_disc is never defined. As written, the prototypes z_ob and z_bg are not class-wise cup/disc prototypes, so it is unclear how the method refines class-specific pseudolabels. This disconnect also makes it hard to see how the contrastive disalignment in Section 3.2 interacts with the refinement stage.","section":"Section 3.1, Eq. (9)"},{"comment":"L_sim is written as a single-pixel cosine similarity with no summation or averaging over pixels; the set of contributing pixels is unspecified. In addition, Algorithm 1 (line 21) writes L_total = L_seg + L_sim without the λ of Eq. (13), and λ is never given in Section 4. The method is therefore not reproducible as stated.","section":"Section 3.2, Eq. (12); Algorithm 1"},{"comment":"This experiment does not ablate either proposed module. It compares divergence metrics (KL, JS) with distance metrics (MMD, Euclidean) on gradient maps, while the proposed loss is a cosine similarity combined with a segmentation loss. No run is reported that removes the gradient-guided prototypes or the contrastive loss. The claim that the 'ablation study validates' the approach is not supported by the presented experiment.","section":"Section 4, 'Ablation Study' / Table 2"},{"comment":"The entire mechanism assumes that source-model Grad-CAM maps localize the optic cup and disc on target images. The paper provides no target-domain validation of this assumption: no Grad-CAM visualizations on target images, no quantitative localization test against target ground-truth masks, and no comparison with source-domain map quality. Table 2 evaluates distributional distances between cup and disc gradient maps, not localization accuracy. If domain shift corrupts these maps, Eq. (9) will discard correct pseudolabels and Eq. (12) will separate wrong features, so the reported gains could arise from extra training signal rather than from gradient-guided refinement. This is a load-bearing point that needs direct evidence.","section":"Section 3.1, Eqs. (4)-(5); Section 4, Table 2"},{"comment":"No error bars or multiple-run statistics are reported. Several improvements over the strongest baselines are small (e.g., cup Dice 80.51 vs 79.78 on RIM-ONE-r3; disc Dice 84.67 vs 83.56 on Drishti-GS). Without variance estimates, the central claim of outperforming state-of-the-art methods is not statistically supported for those comparisons.","section":"Section 4, Table 1"}],"minor_comments":[{"comment":"The denominator uses max(||a||||b||, ε). If the norm product is smaller than ε, the loss is scaled by 1/ε rather than by the actual norm product, which can amplify the loss. Consider using ||a||||b|| + ε or an equivalent standard stabilizer.","section":"Eq. (12)"},{"comment":"The header row is garbled: the repeated 'Cup Disc' entries and the misaligned metric labels make the table difficult to read. The table would benefit from a conventional grouped-header format.","section":"Table 2"},{"comment":"PLPB is described as a 'robust image-classification SFDA network', but the cited work is a fundus image segmentation method; the description should be corrected.","section":"Section 2.2"},{"comment":"The paper gives η=0.05, K=10, and λ is missing. A sensitivity analysis for η and λ would help reproducibility, especially since the refinement mask depends directly on η.","section":"Section 4"},{"comment":"Line 6 uses GAP for the same operation as Eq. (4), but the GAP notation is not defined in the algorithm. Also, the algorithm omits the λ from Eq. (13); this should be aligned.","section":"Algorithm 1"},{"comment":"Figure 1 is placed after Eq. (8) but is not referenced in the surrounding text. Please cite it where the two-stage pipeline is first described.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The core idea has potential merit for medical-image SFDA, and the authors provide an evaluation on two public datasets with code release. However, the manuscript's central mechanism is underspecified and rests on an unvalidated gradient-localization assumption. If the authors can clarify the equations, supply the missing hyperparameter, and provide target-domain validation of the gradient maps, the contribution could become publishable. I see no novelty-disclosure or ethical concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does what it says: it takes DPL-style prototype refinement, adds Grad-CAM-derived gradient maps to weight the prototypes, and throws in a cosine contrastive loss to push cup and disc features apart. The combination is new, the idea is intuitive, and the writing is mostly clear. On the standard REFUGE-to-Drishti-GS/RIM-ONE-r3 benchmarks it beats DPL and PLPB, with the disc Dice gain on RIM-ONE-r3 (90.13 to 94.99) looking real if the numbers are reproducible. The code is promised. That is genuine credit.\n\nThe soft spot is the thing both the reader and I landed on: the whole mechanism assumes the source model's gradient maps reliably localize the optic cup and disc on target images. Eqs. (4)-(5) produce those maps from target data, and then Eqs. (6)-(9) and (11)-(12) lean on them. The paper never checks that this holds. No target-domain Grad-CAM visualizations, no localization accuracy against ground truth, no sanity check that the maps are class-discriminative after domain shift. If the maps are corrupted, the pseudolabel mask drops good pixels and the contrastive loss pushes the wrong features apart, and Table 2 does not resolve this—it compares divergence metrics between the maps, not their localization quality.\n\nThe other issues are more mundane but still need fixing. The so-called ablation (Table 2) does not ablate the proposed components; it compares distance metrics. No error bars in Table 1, no value given for lambda in Eq. (13), and Eq. (12) is underspecified—it doesn't say over which pixels the cosine similarity is summed. The reference for DAE is also wrong (it points to Lee's pseudo-label paper, not a denoising autoencoder work). None of these are fatal if the code reproduces the tables, but they are exactly what a serious reviewer should request.\n\nMy take: this is a legitimate extension of DPL, not a new framework. It will interest people working on source-free medical image segmentation, but it is not a must-read. For a journal, it deserves peer review: the core idea is plausible, the benchmark is standard, and the issues are fixable with a real ablation and target-domain validation. If the reviewers push on the gradient-map assumption, the authors either provide evidence or find that the gains come from something else—either way, that is a useful outcome.","headline":"A sensible, incremental SFDA method for cup/disc segmentation, but the core mechanism is untested on target data and the ablation doesn't ablate.","tokens_in":11813,"tokens_out":2790,"would_cite":false,"duration_ms":32495,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Gradient-guided feature maps refine pseudolabels and disalign optic cup and disc features, enabling source-free domain adaptation for fundus image segmentation.","keywords":["source-free domain adaptation","fundus image segmentation","optic disc and cup","gradient-guided pseudolabels","contrastive learning","cosine similarity","Grad-CAM","glaucoma diagnosis"],"falsifier":"Run Grad-CL with the gradient maps replaced by random or zero masks on the same target datasets; if the performance gain over DPL disappears, the claim that gradient-guided features drive the improvement is confirmed. Alternatively, measure the overlap between the gradient heatmaps and the true cup/disc masks on a heavily shifted target dataset—if the overlap is low, the method's Dice should drop sharply.","tokens_in":10966,"feed_emoji":"👁️","tokens_out":2739,"duration_ms":28706,"temperature":0.7,"pith_summary":"Grad-CL is a source-free domain adaptation method for segmenting the optic cup and disc in fundus images. It starts from a model trained on a labeled source dataset and adapts it to an unlabeled target dataset using two mechanisms: gradient-guided pseudolabel refinement, which uses class-specific gradient maps to sharpen noisy pseudolabels, and a cosine-similarity contrastive loss that pushes cup and disc features apart. The paper reports that this combination outperforms previous unsupervised and source-free adaptation methods on two public fundus datasets, with the largest gains on the harder RIM-ONE-r3 target. If the method holds, it offers a way to adapt medical segmentation models across clinical sites without sharing source data.","feed_headline":"Gradient maps let eye-scan models adapt without source data","feed_subtitle":"Grad-CL refines pseudolabels and disaligns cup/disc features, beating prior SFDA methods on fundus datasets","key_machinery":"Gradient-guided class-specific feature maps (Grad-CAM-style heatmaps e_cup_GC and e_disc_GC) are the load-bearing object. They are computed by global average pooling of gradients of each class logit with respect to the final convolutional feature maps, then combined with raw features to form prototypes (Eq. 6-8) and contrasted via cosine similarity (Eq. 11-12). The gradient maps thus carry the argument: they localize class-discriminative pixels, making pseudolabel refinement and the contrastive disalignment target both class-aware.","core_discovery":"The central claim is that gradient-based saliency maps of the two classes, computed from the pretrained source model on target images, carry information that improves both stages of source-free adaptation. The maps refine pseudolabel reliability by weighting feature prototypes, and they serve as class-specific feature anchors for a contrastive loss that explicitly disaligns optic cup and disc representations. The paper demonstrates the mechanism through ablations comparing divergence versus distance metrics, and through Dice and ASD scores on Drishti-GS and RIM-ONE-r3.","pith_inferences":["The gradient-guided disalignment idea is not restricted to fundus images; any two-class segmentation with a clear object-pair structure could benefit, though the paper does not test this.","Because the gradient maps come from the source model, the method inherits the source model's biases; if the source model localizes classes poorly on target images, errors could be amplified rather than corrected.","An empirical test with larger domain shifts or different backbones (e.g., transformer-based segmentation models) would clarify whether the mechanism is general or specific to DeepLabv3+/MobileNetv2.","A direct check that the gradient maps actually overlap the true cup and disc on target images would validate the core assumption; the paper does not report such localization accuracy."],"forward_implications":["If correct, source-free domain adaptation for fundus segmentation can be improved without any access to source data, easing privacy and confidentiality constraints in clinical settings.","Gradient-guided prototype estimation reduces the influence of noisy pseudolabels compared to prototypes built from combined features, as shown by the improved Dice scores over DPL.","Explicitly disaligning optic cup and disc features reduces misclassification between the two classes and improves boundary delineation, reflected in lower average surface distance.","The method's gains are largest on the more challenging RIM-ONE-r3 target, suggesting it helps when the domain shift is severe.","The approach combines pseudolabeling and contrastive learning in a way that could be adapted to other two-class medical segmentation tasks."],"fun_headline_variants":["Gradient cues sharpen eye-scan adaptation without source data","Gradient maps guide pseudolabels for source-free eye segmentation","Saliency weights clean up optic-disc labels for unlabeled targets","Gradient-guided contrastive learning adapts cup/disc segmentation","Grad-CL: gradient-salient features beat prior source-free adapters"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire pipeline assumes the gradient maps computed from the source model on target images reliably point at the optic cup and disc; if domain shift corrupts this localization, both the pseudolabel refinement and the contrastive objective are built on wrong targets.","fun_headline_variants_meta":{"raw":{"variants":["Gradient cues sharpen eye-scan adaptation without source data","Gradient maps guide pseudolabels for source-free eye segmentation","Saliency weights clean up optic-disc labels for unlabeled targets","Gradient-guided contrastive learning adapts cup/disc segmentation","Grad-CL: gradient-salient features beat prior source-free adapters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001027,"raw_usage":{"total_tokens":4152,"prompt_tokens":714,"completion_tokens":3438,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":458,"completion_tokens_details":{"reasoning_tokens":3350}},"tokens_in":458,"tokens_out":3438,"duration_ms":19527,"temperature":1.0,"reasoning_tokens":3350,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T18:05:23.159491+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Grad-CL with the gradient maps replaced by random or zero masks on the same target datasets; if the performance gain over DPL disappears, the claim that gradient-guided features drive the improvement is confirmed. Alternatively, measure the overlap between the gradient heatmaps and the true cup/disc masks on a heavily shifted target dataset—if the overlap is low, the method's Dice should drop sharply.","supporting_citations":[],"review_version":1}