{"id":"adb560a9-8106-4f15-ae19-605d9812bf37","arxiv_id":"2501.14885","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A prototype-based RBF classifier built on CNN segment embeddings reports 83.02% on ISIC 2016 and 76.15% on ISIC 2017, with explanations traced to cluster prototypes, but the evaluation is compromised by an undefined segment-to-image aggregation rule and self-cited baselines.","lead":"This paper pairs a convolutional neural network with a radial basis function network to sort skin lesion images by comparing small image patches to a set of prototype patches, aiming for accuracy and explainability. It reports 83.02% accuracy on ISIC 2016 and roughly 72 to 76% on ISIC 2017, but the evaluation is undermined by an undefined prediction aggregation step, an abstract/body discrepancy, and self-cited comparison baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported image-level accuracies (83.02%/76.15%) cannot be reproduced from Algorithm 1 because no aggregation rule combines per-segment softmax outputs; the unconstrained dense layer also makes 'most activated prototype' explanations potentially unfaithful to the class decision.","rationale":"The central claim is that the hybrid CNN-RBF framework yields accurate and faithfully interpretable skin cancer classification. For the accuracy part, the method must map a multi-segment image to a single prediction; Algorithm 1 stops at per-segment softmax outputs. The text's phrase 'averages decisions across localized patches' is not an algorithm. Because no aggregation rule is specified, the reported 83.02% and 76.15% numbers cannot be regenerated from the described method, and different plausible aggregation rules will generally yield different accuracies. This is not a stylistic omission: it is the missing link between the model and its headline numbers. Without it, the evaluation protocol is undefined, which alone justifies rejection. The secondary issue—that Step 10's 'most activated prototype' may not correspond to the softmax decision because W is unconstrained—further undermines the interpretability claim, but the aggregation gap is sufficient to block verification of the central empirical claim. I agree with the reader's weakest assumption and recommend no change to the REJECT verdict, while noting that the paper could be revised by adding an explicit aggregation formula, aligning abstract/body numbers, and releasing code.","tokens_in":10757,"tokens_out":5027,"duration_ms":46299,"concrete_test":"Implement Algorithm 1 exactly as written and compute image-level accuracy on ISIC 2016 and ISIC 2017 under each plausible aggregation rule: (a) average of segment softmax probabilities, (b) majority vote over segment argmax labels, and (c) max-confidence segment label. Compare the three resulting accuracies to the claimed 83.02% and 76.15%. If one rule reproduces the reported numbers, that rule must be stated in the paper; if none does, the headline accuracy is not reproducible from the described method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline empirical claim is image-level classification accuracy on ISIC 2016 and ISIC 2017 (Table 5: 83.02% and 76.15% with ResNet-50). Algorithm 1 computes per-segment embeddings Z_i, per-segment RBF activations, and per-segment softmax probabilities p(y=c|z) (Steps 4–6 and 9), but never defines how the n=4 segment predictions are combined into an image-level prediction. The introduction vaguely says the approach 'averages decisions across localized patches,' but no formula, majority-vote rule, or pooling operation appears in the algorithm or evaluation section. Consequently, the reported accuracies are not well-defined: treating the four segments as independent test samples would yield a segment-level accuracy that need not equal any image-level accuracy; averaging softmax outputs, majority voting, or max-confidence selection would each give different numbers. Without the aggregation rule, the central quantitative claim is unverifiable and potentially misrepresents the method. A second, compounding flaw: Step 10 defines the explanation as the most activated prototype argmax_i φ_i(z), but Step 6 classifies via softmax(W·Φ(z)+b) with an unconstrained dense layer. The predicted class is the argmax of a learned weighted sum over all prototypes, not necessarily the class of the maximally activated prototype; thus the explanation is not guaranteed to be faithful to the decision. Both issues are load-bearing for the paper's dual claims of accuracy and interpretability.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid interpretable framework for skin cancer diagnosis that combines pre-trained CNN embeddings (VGG-16, ResNet-50) with a Radial Basis Function network. Each image is segmented with SLIC, segment embeddings are clustered to form prototypes, and an RBF activation vector is fed to a dense softmax layer for classification; the explanation is defined as the most activated prototype per segment. The method is evaluated on ISIC 2016 and ISIC 2017, with the main reported results being 83.02% accuracy on ISIC 2016 and 76.15% on ISIC 2017 using ResNet-50. The paper claims that this architecture provides both competitive accuracy and faithful, prototype-based interpretability.","tokens_in":41,"tokens_out":3770,"duration_ms":129579,"significance":"If the central claims were substantiated, the paper would address a real need: an interpretable, prototype-based diagnostic model for skin cancer with localized explanations that clinicians can trace. The motivation for combining segmentation, prototype selection, and RBF networks is clear, and the authors explicitly aim for ante-hoc interpretability rather than post-hoc approximation. The paper also attempts to compare against ProtoPNet and ProtoPShare, which is appropriate. However, the current manuscript does not provide a verifiable central claim: the image-level accuracy is not reproducible because the segment-to-image aggregation step is undefined, the reported ISIC 2017 accuracy is inconsistent between the abstract and the results section, and the baseline comparisons contain circular citations. The interpretability claim is also not guaranteed by the architecture. These issues are load-bearing for both the accuracy and the interpretability contributions.","major_comments":[{"comment":"The abstract reports 72.15% accuracy on ISIC 2017 with ResNet-50, while Section 4 and Table 5 report 76.15% for the same configuration, and the conclusion also says 76.15%. The manuscript never acknowledges or reconciles this discrepancy, so the central empirical claim is ambiguous and the reader cannot tell which number is the reported result.","section":"Abstract vs. Section 4 and Table 5"},{"comment":"The algorithm operates per segment: Step 1 segments the image into n regions, Steps 2-6 compute embeddings, RBF activations, and softmax probabilities for each segment z, and Step 9 gives y_hat = argmax_c p(y=c|z). No step combines the n segment predictions, segment embeddings, or segment probabilities into an image-level prediction. The introduction's statement that the approach 'averages decisions across localized patches' is never formalized, so the reported image-level accuracies of 83.02% and 76.15% are not well-defined and cannot be reproduced.","section":"Table 1, Algorithm Steps 1-9"},{"comment":"The comparison baselines ProtoPNet and ProtoPShare are cited to reference [44], which is the authors' own CA-SoftNet paper, rather than to the original ProtoPNet [40] and ProtoPShare [41] sources. This misattributes the baselines and creates a circular comparison in which the proposed model is evaluated against numbers that are not attributed to the actual methods. Additionally, the row 'ResNet-50 [1]' cites reference [1], which is a systematic review, not the ResNet-50 architecture, and the ProtoPNet/ProtoPShare entries omit the '%' sign, making the table internally inconsistent.","section":"Table 5"},{"comment":"Section 3.1.6 states that prototype selection is based on K-Means clustering, while Algorithm 1 (Table 1), Table 3, and the discussion in Section 4 all refer to K-Medoids. Since prototype selection is the mechanism claimed to provide interpretability, the manuscript must specify which clustering algorithm is actually used; the current inconsistency prevents replication.","section":"Section 3.1.6 vs. Table 1 and Table 3"},{"comment":"Classification is performed by softmax(W·Phi(z)+b) over an unconstrained dense layer, but the explanation is defined as argmax_i phi_i(z), the most activated prototype. Since the learned weight matrix W can reweight prototype activations arbitrarily, the predicted class is not necessarily the class of the most activated prototype. Therefore the explanation is not guaranteed to be faithful to the model's decision, contradicting the paper's claim that explanations are generated through the exact mechanism within the model.","section":"Table 1, Steps 6 and 10"}],"minor_comments":[{"comment":"The affiliation contains a typo: 'Paksitan' should be 'Pakistan'.","section":"Title page"},{"comment":"The table lists 'PyTorch (version 3.12)'; 3.12 is a Python version, not a PyTorch version, and the actual PyTorch version is not given.","section":"Section 3.1.1, Table 2"},{"comment":"The text says 'compactness=100 and sigma=1.*3'; this appears to be a formatting error for sigma=1.3 and should be clarified.","section":"Section 3.1.3"},{"comment":"The ProtoPNet and ProtoPShare rows omit the '%' sign, and the table mixes numbers with and without percentage symbols, making the comparison difficult to read.","section":"Table 5"},{"comment":"The focal loss equation has a typo in the label ('Lf ocal') and uses notation inconsistently with the loss expression in Table 1; the subscripts and arguments of the loss terms should be aligned.","section":"Equation (5)"},{"comment":"Several in-text citations do not match the reference list, including [1] used for ResNet-50 in Table 5, [2] used for GDPR in the introduction, and the placement of [44] for ProtoPNet/ProtoPShare; the reference numbering and attribution should be corrected throughout.","section":"References"}],"recommendation":"reject","confidential_remarks":"The pattern of citations in Table 5 is concerning: baseline numbers for ProtoPNet and ProtoPShare are attributed to the authors' own CA-SoftNet paper rather than to the original sources, which goes beyond a simple typo and undermines the validity of the comparison. I would ask the editor to verify the provenance of all baseline numbers and to check whether the authors have disclosed the relationship between this manuscript and their prior work [43, 44]."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the idea is sensible but the central empirical claim is not defined by the paper's own algorithm. The reader's take is on target, and the stress-test note holds up completely.\n\nWhat's actually new: a concrete pipeline—SLIC superpixels, frozen CNN embeddings, K-medoids prototype selection, RBF similarity layer—applied to ISIC 2016 and 2017. That specific combination isn't exactly in the prior literature. The paper also does a few things well: the motivation for prototype-based interpretability in skin cancer is clearly argued, the silhouette scores give a quick sanity check on cluster quality, and the related work on ProtoPNet, ProtoPShare, and ProtoConcepts is reasonably accurate apart from the citation issues I'll mention.\n\nThe soft spots are load-bearing. First, Algorithm 1 computes per-segment embeddings and per-segment softmax predictions, but never defines how the four segments become one image-level prediction. The intro says the model 'averages decisions across localized patches,' but there's no formula, no majority vote, no pooling. Without that step, the reported 83.02% and 76.15% accuracies are undefined. Second, the explanation mechanism (most activated prototype) is not guaranteed to match the classification decision, because classification goes through an unconstrained dense layer over all RBF activations. The paper does not address this disconnect.\n\nThere's also an internal inconsistency: the abstract reports 72.15% for ISIC 2017 with ResNet50, while Section 4 and Table 5 report 76.15%. And Table 5 cites the ProtoPNet and ProtoPShare baselines to the authors' own CA-SoftNet paper [44] instead of the original sources [40] and [41]; that's circular and makes those baseline numbers unverifiable. No code, data, or error bars are provided.\n\nThese are fixable: specify the aggregation rule, align the numbers, cite external baselines, and release code and hyperparameters. But as written, the paper's central claims can't be independently reproduced.\n\nWho is this for? Readers interested in prototype-based interpretability might want to see the framework, but only after the method is properly specified. A serious referee could ask for the missing details; the paper deserves review rather than desk rejection because the combination, while not highly novel, is plausible and the application is relevant. My recommendation: engage with it, ask for the missing aggregation rule and corrected baselines, and don't trust the accuracy numbers until then.","headline":"The paper's central accuracy claim is undefined because Algorithm 1 never defines how segment-level predictions become image-level predictions, and the abstract's 72.15% conflicts with the body's 76.15% for ISIC 2017.","tokens_in":11650,"tokens_out":2654,"would_cite":false,"duration_ms":27783,"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":"The paper claims that a hybrid CNN-RBF network can classify skin lesions with competitive accuracy while making each prediction traceable to a specific image prototype, reporting 83.02% on ISIC 2016 and 76.15% on ISIC 2017 using ResNet-50…","keywords":["skin cancer diagnosis","radial basis function networks","explainable AI","prototype-based classification","hybrid deep learning","K-medoids clustering","ISIC dataset","interpretable machine learning"],"falsifier":"Run the pipeline and aggregate the four per-segment predictions by majority vote, by averaged softmax, and by maximum softmax; if none of these reproduces the reported 83.02% (ISIC 2016) and 76.15% (ISIC 2017) figures, the claimed accuracies are not defined by the algorithm in the paper.","tokens_in":10499,"feed_emoji":"🩺","tokens_out":6926,"duration_ms":51153,"temperature":0.7,"pith_summary":"The paper proposes a hybrid model for skin-cancer diagnosis that splits each image into segments, embeds each segment with a pretrained CNN, and classifies the segment embeddings with an RBF network whose neurons are learned prototypes. The central claim is that this design reaches competitive accuracy—83.02% on ISIC 2016 and 76.15% on ISIC 2017 with ResNet-50 embeddings—while making each prediction explainable by pointing to the prototype that fired most strongly. The authors argue this prototype-based explanation is faithful because it uses the model's actual decision mechanism rather than a post-hoc approximation. A sympathetic reader would care because it suggests diagnostic AI can be both accurate and transparent enough for clinical use.","feed_headline":"Hybrid model reads skin lesions in segments, hits 83% and shows why","feed_subtitle":"Each prediction points to the most similar prototype image segment, giving clinicians visible evidence for the call.","key_machinery":"The central mechanism is the RBF activation layer: each segment embedding $z$ produces a vector $\\Phi(z)$ whose entries are Gaussian similarities $\\phi_i(z) = \\exp(-\\|z-c_i\\|^2 / 2\\sigma^2)$ to $k$ cluster prototypes $c_i$ chosen by K-medoids clustering over user-selected salient segments. A softmax over a dense layer fed by $\\Phi(z)$ yields class probabilities, and the argmax over $\\phi_i(z)$ identifies the prototype that explains the decision. The prototypes themselves are shown as the corresponding image segments, grounding each prediction in a visible patch of skin.","core_discovery":"The paper's core claim is that a hybrid classifier—pretrained CNN embeddings followed by an RBF network over K-medoid prototypes—can classify skin lesions with accuracy competitive with standard deep models while remaining transparent. Concretely, the paper reports 83.02% accuracy on ISIC 2016 and 76.15% on ISIC 2017 using ResNet-50 embeddings, with VGG-16 embeddings giving 79.54% and 70.09%. The new ingredient is that the classifier's decision for each image segment is a softmax over Gaussian similarities to class prototypes, so the explanation for a prediction is literally the prototype segment that activated most strongly. The paper argues that because the explanation comes from the same RBF computation that produces the prediction, it is faithful by construction, unlike post-hoc methods such as Grad-CAM.","pith_inferences":["If the missing aggregation rule is specified as a simple majority vote over the four segments, the reported accuracy could be checked directly; until then the headline numbers are not reproducible from the algorithm as written.","The prototype explanation inherits the quality of the clustering step: if K-medoids picks a prototype that is visually similar but clinically irrelevant, the explanation would look plausible while pointing at the wrong feature.","A natural next experiment would compare explanation faithfulness across RBF prototypes, ProtoPNet-style prototypes, and Grad-CAM on the same backbone, measuring whether clinicians rank the prototype explanations as more useful.","One could test the framework on a class-imbalanced real-world dataset without SMOTE or focal loss, to isolate how much of the reported accuracy comes from the RBF architecture versus the augmentation and loss tricks."],"forward_implications":["On ISIC 2016 and ISIC 2017, the RBF-NN with ResNet-50 embeddings reaches 83.02% and 76.15% accuracy, beating its VGG-16 variant while remaining less accurate than the non-interpretable ResNet-50 baseline reported at 92.90% and 83.00%.","Each prediction is explained by the single prototype segment whose RBF activation is strongest, so the explanation is an actual image region rather than a heatmap approximation.","Using deeper embeddings (ResNet-50 over VGG-16) improves both clustering coherence, with silhouette scores rising from 0.63 to 0.72 on ISIC 2016, and final classification accuracy.","The authors claim the same architecture extends to other medical imaging tasks, such as histopathology or radiology, where localized, transparent evidence is needed."],"supporting_citations":[{"why":"ISIC 2016 dataset source; provides the training and test images for the first benchmark.","marker":"[12]"},{"why":"ISIC 2017 dataset source; provides the three-class training, validation, and test split.","marker":"[13]"},{"why":"ResNet-50 definition; the backbone that yields the paper's best embeddings and accuracy.","marker":"[24]"},{"why":"VGG-16 definition; the shallower backbone used for the alternative embeddings.","marker":"[26]"},{"why":"Radial basis function networks; supplies the RBF activation mechanism at the core of the classifier.","marker":"[38]"},{"why":"ProtoPNet; the prototype-based interpretable model the paper contrasts with and compares against.","marker":"[40]"},{"why":"PatchNet-21; an interpretable baseline included in the accuracy comparison table.","marker":"[45]"}],"fun_headline_variants":["Skin cancer AI that shows its work via prototype segments","Hybrid model explains each skin lesion call with a prototype","RBF network reads lesions in parts, hits 83% and explains why","AI for skin cancer that points to the exact image segment it used","Interpretable deep learning: skin cancer calls traced to prototypes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported accuracies depend on an unstated rule for converting the model's per-segment predictions into a single image-level prediction, because the algorithm as written never combines the $n$ segments of an image.","fun_headline_variants_meta":{"raw":{"variants":["Skin cancer AI that shows its work via prototype segments","Hybrid model explains each skin lesion call with a prototype","RBF network reads lesions in parts, hits 83% and explains why","AI for skin cancer that points to the exact image segment it used","Interpretable deep learning: skin cancer calls traced to prototypes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000178,"raw_usage":{"total_tokens":1323,"prompt_tokens":1000,"completion_tokens":323,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":616,"completion_tokens_details":{"reasoning_tokens":237}},"tokens_in":616,"tokens_out":323,"duration_ms":4148,"temperature":1.0,"reasoning_tokens":237,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:49:23.879953+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline and aggregate the four per-segment predictions by majority vote, by averaged softmax, and by maximum softmax; if none of these reproduces the reported 83.02% (ISIC 2016) and 76.15% (ISIC 2017) figures, the claimed accuracies are not defined by the algorithm in the paper.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ISIC 2017 dataset source; provides the three-class training, validation, and test split."},{"cited_title":"Patchnet: Interpretable Neural Networks for Image Classification","cited_arxiv_id":"1705.08078","evidence_quote":"PatchNet-21; an interpretable baseline included in the accuracy comparison table."}],"review_version":1}