{"id":"1167158e-1be3-4f5e-a732-67b009dfa226","arxiv_id":"2507.23709","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Averaging pixel attribution maps over multiple MC Dropout forward passes, plus a coefficient-of-variation risk map, raises ADCC scores over single-pass baselines on pCLE and ImageNet.","lead":"This paper combines Monte Carlo Dropout with pixel attribution maps: the same classifier runs many times with random neurons switched off, the explanation maps are averaged, and a risk map shows which pixels vary most. The aim is to make AI explanations for brain tumour tissue images less overconfident and easier for surgeons to trust.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No classification accuracy is reported under MC Dropout; if accuracy degrades, the improved ADCC and risk maps do not support clinical use.","rationale":"The reader's weakest assumption is exactly the one I consider most load-bearing: the paper does not validate that the classifier remains accurate when Dropout is enabled at test time, even though all explainability claims are computed under that regime. Without this check, the entire framework could be explaining a degraded model, which would make the improved ADCC scores misleading for surgical decision support. I agree with the reader's assessment that the evaluation is insufficient, and the fix is straightforward to request. I also noticed internal numerical inconsistencies (Recipro-CAM ADCC differs between the main table and the supplement; Score-CAM proposed results are identical to original), which reinforce the need for a careful accuracy and consistency audit. However, these do not constitute proof of failure; they are missing or unreliable evidence. Therefore the appropriate outcome is a conditional acceptance pending the accuracy verification, exactly as the reader recommended. No change to the verdict is needed.","tokens_in":167,"tokens_out":4638,"duration_ms":60038,"concrete_test":"On the same pCLE test split used for Tables 1 and 3, compute top-1 classification accuracy (and, if possible, expected calibration error) for deterministic inference and for T=10 MC Dropout with softmax averaging, for both ResNet18 and MobileNetV2. If the MC Dropout accuracy is materially lower than the deterministic accuracy (e.g., a drop of more than 5 percentage points), the clinical explainability claim fails; if accuracy is comparable, the concern is resolved and the ADCC improvements can be taken at face value.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that stochastic averaging of T=10 MC Dropout pixel attribution maps yields more trustworthy explanations for tissue characterisation. This presumes the model's output logits remain reliable when Dropout is enabled at test time. The paper never reports classification accuracy, precision, recall, or calibration for the MC Dropout regime. In the supplementary tables, several baseline models already show large average drop values on the pCLE data (e.g., MobileNetV2 with Grad-CAM has AD of 73.8), and the proposed method reduces this to 59.3, but the models' overall ability to classify the three tumour types is never quantified. If MC Dropout degrades classification performance, the enhanced PA maps and CV risk maps are built on a less competent classifier, and the reported ADCC improvements may reflect a smoother but less clinically useful explanation. Furthermore, the ADCC metric uses the model's own output scores to compute coherency and average drop; with stochastic forward passes, these scores are noisy, and the paper does not specify whether ADCC is computed using averaged logits or a single forward pass. A separate but reinforcing issue is internal inconsistency in the reported numbers: Recipro-CAM's original ADCC for ResNet18 is 66.5 in Table 1 but 72.8 in Supplementary Table 3, and Score-CAM's proposed results are identical to its original results, suggesting the MC Dropout procedure may not have been applied consistently. These gaps leave the core claim unsupported as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a stochastic pixel-attribution framework: at test time a CNN trained with Dropout is run T times with Dropout enabled, a PA method computes a map per pass, and the T maps are averaged pixel-wise to form an enhanced PA map; the coefficient of variation across the T maps is presented as a pixel-wise risk estimate. The method is evaluated with Grad-CAM, Grad-CAM++, SmoothGrad-CAM++, Score-CAM, and Recipro-CAM on a pCLE brain-tumour dataset (ResNet18, MobileNetV2) and on ImageNet (ResNet50), using ADCC and latency. The authors claim that the framework improves on all tested SOTA PA methods' ADCC and that the CV map provides trustworthy risk information for intraoperative tissue characterisation.","tokens_in":9696,"tokens_out":5760,"duration_ms":57451,"significance":"The idea is simple, model-agnostic, and PA-method-agnostic, and the CV-based risk map is a genuinely useful addition for clinical explainability: it lets a surgeon see not only which pixels are salient but also how stable that saliency is across stochastic passes. The evaluation spans both medical and natural-image domains, and the qualitative figures support the plausibility of the risk visualisation. If the missing controls and consistency issues are resolved, this would be a solid contribution to intraoperative explainability. In its current form, however, the central quantitative claim is not yet established because the comparison changes two variables at once and the stochastic model's classification competence is never reported.","major_comments":[{"comment":"The central comparison changes two variables at once: test-time Dropout is enabled and T=10 PA maps are averaged, while the baselines use neither. No experiment isolates the averaging effect from the Dropout effect, and no classification accuracy, F1, or calibration is reported for the MC Dropout regime. If enabling Dropout degrades the classifier, then the improved ADCC values and the qualitative PA maps are produced by a less competent model, and the 'reduced overconfidence' claim is not supported. Please add an ablation with (i) T=1 with Dropout, (ii) T=10 without Dropout, and (iii) the original deterministic model, and report accuracy/calibration for all regimes.","section":"§3.0.2, §3.1.1, Table 1"},{"comment":"T is selected on the same ADCC metric and the same benchmark used for the final evaluation, and the reported ADCC differences are often small (e.g., ResNet18 Grad-CAM 76.6 to 77.7; MobileNetV2 Grad-CAM 48.0 vs 47.9 at nearby T values in Table 5). No error bars, standard deviations, or patient-level repeated evaluations are provided. This selection procedure can inflate the claimed gain. Please use a held-out validation split to choose T and report variability over runs or patients.","section":"§3.1, Table 5"},{"comment":"There are internal inconsistencies that must be resolved. Recipro-CAM's original ADCC for ResNet18 is 66.5 in Table 1 but 72.8 in Supplementary Table 3; MobileNetV2 Recipro-CAM is 33.3 vs 35.8; and original Grad-CAM latency for ResNet18 is 75.6 ms in Table 1 but 5.6 ms in Supplementary Table 3. In addition, the 'Proposed' Score-CAM rows are numerically identical to the 'Original' rows on all three datasets, even though the latency scales roughly by a factor of 10. Please state whether ADCC is computed with averaged logits or with a single forward pass, explain the Score-CAM result, and correct the inconsistent values.","section":"Table 1 vs Supplementary Table 3; Table 2 vs Supplementary Table 4"},{"comment":"The ADCC definition in Eq. (4) is written as the reciprocal of a sum of reciprocals, which omits the factor 3 required for a harmonic mean of three components as defined in the cited work [26]. Moreover, Eq. (6) defines Complexity as 100 times the L1 norm, and Eq. (4) then uses 1 - Complexity in a denominator, which is only meaningful if Complexity is normalised to [0,1]. Please verify the formula against [26] and clarify the normalisation so that the reported ADCC values are comparable with the literature.","section":"§3.0.3, Eq. (4)"}],"minor_comments":[{"comment":"The text says 'Our proposed method improves ADDC' but the metric is ADCC; please fix the typo.","section":"§2"},{"comment":"The variance formula in Eq. (1) uses a transpose notation on what appear to be scalar pixel attribution values; please clarify whether this is a vectorised expression over pixels or a typographical error.","section":"§2, Eq. (1)"},{"comment":"The implementation paragraph first gives separate learning rates for ResNet18 (0.001) and MobileNetV2 (0.01), then says 'Both were trained using Adam-W ... with a learning rate of 0.001'. Please disambiguate which optimiser/learning-rate combination was used for each model.","section":"§3.0.2"},{"comment":"There are several typos and inconsistencies: 'MoblieNetV2' (Supplementary Table 3), 'spurrious', 'adverserial', 'pixe-wise', 'classfication', 'Dropout it not enabled', and 'assitive DL model'. A careful proofread is needed.","section":"Throughout"},{"comment":"The claim of being 'the first approach which incorporates risk estimation into a pixel attribution method' is strong; earlier uncertainty-aware saliency works should be discussed, or the claim should be tempered to 'first in the context of intraoperative tissue characterisation'.","section":"Abstract and Introduction"}],"recommendation":"major_revision","confidential_remarks":"I share the reader's main concern: the missing classification accuracy under MC Dropout is the key blocker, because the ADCC improvements may be generated on a degraded classifier. The internal numerical inconsistencies reinforce the need for a careful revision before the paper can be considered for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a simple, sensible idea that mostly delivers what it promises, but the evaluation is not tight enough to support the headline as written.\n\nWhat's actually new: the combination of MC Dropout with pixel attribution—run T stochastic forward passes, average the PA maps, and report per-pixel coefficient of variation as a risk map. That's a cheap, model-agnostic wrapper, and I don't see it in the cited literature. The qualitative examples in Figures 2, 5, and 6 show why a surgeon would find the CV map useful: you can see which highlighted regions are stable across passes. The ADCC gains over deterministic Grad-CAM, Grad-CAM++, and Recipro-CAM are consistent across pCLE and ImageNet, and the latency tables are honest. The T-sweep in Supplementary C is also there. Credit where due: it's a useful practical contribution.\n\nSoft spots, in order of importance. First, no classification accuracy is reported with dropout enabled at test time. The central assumption is that the stochastic model is still a competent classifier. If dropout degrades accuracy, the improved ADCC numbers describe a worse model, and the \"trustworthy explanation\" story collapses. Second, no ablation separates MC Dropout from averaging. T=10 averaging alone might explain most of the gain; the method should compare deterministic T=10 averaging against stochastic T=10 averaging with dropout. Third, T is tuned on the same ADCC metric used for the final comparison. That is selection on the evaluation set, and while the effect looks small, it should be reported on a validation split or shown to be flat. Fourth, the tables are internally inconsistent: Recipro-CAM's original ResNet18 ADCC is 66.5 in Table 1 but 72.8 in Supplementary Table 3, and Score-CAM's proposed ADCC is identical to its original in every row, which makes me wonder whether the wrapper was actually applied to Score-CAM. The identical ADCC is suspicious even if the mechanism differs. Fifth, no error bars or significance tests; with what appears to be a single seed and a private pCLE dataset, small ADCC differences are not reliable.\n\nNone of this kills the idea. The method is plausible, it is built on public tools, and the risk-map visualization is a genuine value-add. It just needs a cleaner evaluation before I would want it cited as SOTA in a clinical paper.\n\nRecommendation: send it out for review, but the dropout-accuracy ablation, the deterministic-vs-stochastic averaging ablation, and the table consistency checks should be must-fix items before acceptance. For a reading group, it is a good case study in how easy it is to make explainability metrics look conclusive when the underlying model's competence is not checked.","headline":"A cheap, sensible wrapper that puts MC Dropout inside pixel attribution and adds a CV risk map; the idea is useful but the evaluation needs an accuracy check, an ablation, and table cleanup before the headline claims hold.","tokens_in":10253,"tokens_out":1951,"would_cite":false,"duration_ms":22089,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Repeated stochastic forward passes with dropout sharpen pixel-attribution maps and add a per-pixel risk score for tissue-characterisation images.","keywords":["Explainability","Pixel attribution maps","MC Dropout","Risk estimation","Coefficient of variation","Tissue characterisation","Confocal laser endomicroscopy","ADCC"],"falsifier":"Measure top-1 classification accuracy and calibration on the held-out pCLE and ImageNet test sets under the proposed protocol—dropout enabled, T=10 forward passes. If accuracy falls materially below the deterministic model's accuracy, then the higher ADCC scores would be describing a worse classifier, and the claim that the enhanced heatmap is a better explanation for clinical use would not stand.","tokens_in":9237,"feed_emoji":"🧠","tokens_out":14596,"duration_ms":133343,"temperature":0.7,"pith_summary":"The paper proposes a way to make explainability maps for tissue-classification AI less overconfident and more usable in surgery. It runs a dropout-trained classifier ten times with dropout enabled at test time, computes a pixel-attribution (saliency) heatmap for each run, and averages the ten heatmaps into a single enhanced map; the run-to-run spread of each pixel, divided by its average, becomes a pixel-wise risk estimate. The paper reports that this stochastic averaging improves the ADCC metric (a harmonic mean of average drop, coherency, and complexity) relative to the same heatmap methods run deterministically on brain-tumour images from probe-based confocal laser endomicroscopy (pCLE) and on ImageNet, and that the risk maps line up with the salient regions. A surgeon would get, in one output, both a sharper explanation and a signal about which parts of that explanation to trust.","feed_headline":"Ten dropout runs sharpen tumor heatmaps and add a risk map","feed_subtitle":"Repeated dropout passes produce a sharper saliency map and flag the pixels the surgeon should not trust.","key_machinery":"The central object is a PA volume: T stochastic forward passes of the classification model with dropout enabled, each converted into a pixel-attribution map by the same PA method and stacked along a third dimension. From this volume the method derives two outputs: the enhanced PA map given by the per-pixel expectation over the T maps, and a risk map given by the coefficient of variation $\\mathrm{CV}_{i,j} = \\operatorname{std}(S_{i,j})/\\operatorname{E}(S_{i,j})$. The coefficient of variation is the load-bearing choice: because pixels correspond to different semantic features with different attribution scales, raw variance would be misleading, while dividing by the pixel's mean makes the risk scores comparable. The mechanism converts dropout's cheap, repeated model averaging into both denoising (the expectation) and disagreement measurement (the CV).","core_discovery":"The paper's central claim is that a single deterministic pixel-attribution map inherits the overconfidence of the neural network it explains, and that this can be corrected by treating attribution as a stochastic measurement. With dropout enabled at test time, the classifier is run T times (T=10 in the experiments) and any pixel-attribution method—Grad-CAM, Grad-CAM++, SmoothGrad-CAM++, Score-CAM, or Recipro-CAM—is recomputed on each run, producing a volume of maps. The enhanced map is the pixel-wise expectation $E(S_{i,j}) = \\frac{1}{T}\\sum_{t=1}^{T} f_s(\\hat{Y}_t)_{i,j}$, and the risk of that map is the per-pixel coefficient of variation $\\mathrm{CV}_{i,j}=\\sqrt{\\mathrm{Var}(S_{i,j})}/E(S_{i,j})$, chosen because it allows comparison across pixels with different scales. On pCLE data and ImageNet this raises ADCC relative to the deterministic versions of the same methods and ties Score-CAM, while adding a risk surface that concentrates low risk where attribution is high. The paper concludes that this two-channel output is a more trustworthy explainability tool for intraoperative tissue characterisation.","pith_inferences":["A natural extension, not tested in the paper, would threshold the CV map and automatically grey out high-risk pixels, turning the risk estimate from a passive visualisation into an active safeguard against acting on untrustworthy highlights.","We infer the same PA volume could double as an out-of-distribution detector: dropout disagreement tends to rise on inputs unlike the training data, so unusually high global CV could flag images the classifier was never competent to explain.","A direct test of the trustworthiness claim would compare low-CV pixels against regions a clinician independently marks as diagnostically relevant, or against known spurious-correlation artefacts such as the crack visible in the meningioma videos shown in the paper's supplement."],"forward_implications":["When surgeons see a bright but high-CV region, the method tells them not to treat that highlight as evidence, even though a single deterministic map would present it confidently.","Because the procedure is agnostic to the base PA method, any existing heatmap technique can be upgraded to the stochastic-averaging version without retraining the classifier.","The T-fold increase in latency, roughly ten attribution computations per image, is the main deployment cost, and the paper selects T=10 as the trade-off point between ADCC gain and speed.","The ImageNet results suggest the mechanism transfers beyond medical microscopy, so the same heatmap-plus-risk output should apply to other image classification settings."],"supporting_citations":[{"why":"Establishes MC Dropout: test-time forward passes with dropout approximate model uncertainty and provide cheap model averaging, the source of the PA volume.","marker":"[13]"},{"why":"Defines the ADCC metric used to compare the enhanced PA maps against deterministic baselines.","marker":"[26]"},{"why":"Grad-CAM is one of the PA methods the stochastic volume is applied to and evaluated against.","marker":"[29]"},{"why":"Grad-CAM++ is a baseline PA method whose deterministic ADCC is improved by the proposed averaging.","marker":"[6]"},{"why":"Recipro-CAM is the fast gradient-free PA method whose combination with the proposed method produces large ADCC gains.","marker":"[5]"},{"why":"Score-CAM is the strongest comparison PA method; the proposed method ties its ADCC while adding risk estimation.","marker":"[31]"},{"why":"ImageNet is the natural-image benchmark that demonstrates the method generalises beyond pCLE medical data.","marker":"[8]"},{"why":"ResNet-18 and ResNet-50 are the classifier architectures trained with dropout and run stochastically at test time.","marker":"[17]"}],"fun_headline_variants":["Dropout runs turn confident heatmaps into honest risk maps","Shake the classifier to spot unreliable pixels in saliency maps","Monte Carlo dropout adds risk to pixel attribution maps","Stop trusting overconfident heatmaps: get a risk map","Repeated dropout reveals when saliency maps lie"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the network still classifies accurately, and that its prediction scores remain meaningful, when dropout is switched on at test time; the paper reports explanation metrics but never reports classification accuracy in that mode.","fun_headline_variants_meta":{"raw":{"variants":["Dropout runs turn confident heatmaps into honest risk maps","Shake the classifier to spot unreliable pixels in saliency maps","Monte Carlo dropout adds risk to pixel attribution maps","Stop trusting overconfident heatmaps: get a risk map","Repeated dropout reveals when saliency maps lie"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00043,"raw_usage":{"total_tokens":2224,"prompt_tokens":1000,"completion_tokens":1224,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":616,"completion_tokens_details":{"reasoning_tokens":1144}},"tokens_in":616,"tokens_out":1224,"duration_ms":11009,"temperature":1.0,"reasoning_tokens":1144,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:26:52.600739+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure top-1 classification accuracy and calibration on the held-out pCLE and ImageNet test sets under the proposed protocol—dropout enabled, T=10 forward passes. If accuracy falls materially below the deterministic model's accuracy, then the higher ADCC scores would be describing a worse classifier, and the claim that the enhanced heatmap is a better explanation for clinical use would not stand.","supporting_citations":[{"cited_title":"Dropout as a Bayesian Approximation: Represent- ing Model Uncertainty in Deep Learning","cited_arxiv_id":null,"evidence_quote":"Establishes MC Dropout: test-time forward passes with dropout approximate model uncertainty and provide cheap model averaging, the source of the PA volume."},{"cited_title":"Revisiting The Evaluation of Class Activation Mapping for Explainability: A Novel Metric and Exper- imental Analysis","cited_arxiv_id":null,"evidence_quote":"Defines the ADCC metric used to compare the enhanced PA maps against deterministic baselines."},{"cited_title":"Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra","cited_arxiv_id":null,"evidence_quote":"Grad-CAM is one of the PA methods the stochastic volume is applied to and evaluated against."},{"cited_title":"Grad-CAM++: Improved Visual Explanations for Deep Convolutional Net- works","cited_arxiv_id":null,"evidence_quote":"Grad-CAM++ is a baseline PA method whose deterministic ADCC is improved by the proposed averaging."},{"cited_title":"Recipro-CAM:Gradient-freereciprocalclassactivation map","cited_arxiv_id":null,"evidence_quote":"Recipro-CAM is the fast gradient-free PA method whose combination with the proposed method produces large ADCC gains."},{"cited_title":"Score-CAM: Score-Weighted Visual Explanations for Convolu- tional Neural Networks","cited_arxiv_id":null,"evidence_quote":"Score-CAM is the strongest comparison PA method; the proposed method ties its ADCC while adding risk estimation."},{"cited_title":"Imagenet: A large-scale hierarchical image database","cited_arxiv_id":null,"evidence_quote":"ImageNet is the natural-image benchmark that demonstrates the method generalises beyond pCLE medical data."},{"cited_title":"Deep Residual Learning for Image Recognition","cited_arxiv_id":null,"evidence_quote":"ResNet-18 and ResNet-50 are the classifier architectures trained with dropout and run stochastically at test time."}],"review_version":1}