{"id":"9c39fd7c-79f5-40c5-9f77-e7b80845a909","arxiv_id":"2412.01590","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"NCDD subtracts a test image's distance to the closest class centroid from its summed distances to the other centroids, with norm-based weights, and this score improves out-of-distribution detection for GI endoscopy abnormalities in the reported benchmarks.","lead":"An endoscopy AI trained only to recognize healthy landmarks gets a new scoring rule, NCDD, that flags abnormal images by comparing distances to the nearest healthy-feature centroid and to the other centroids. The authors report this rule detects GI diseases better than several established out-of-distribution detectors on two public datasets across four neural-network backbones.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Hyperparameters α1/α2 are tuned on synthetic OOD (random rectangles + speckle), not on real disease classes; if that proxy is unrepresentative, the reported NCDD gains are not established.","rationale":"The reader's weakest assumption — that the synthetic validation OOD used to tune α1 and α2 may not behave like real GI abnormalities — is indeed the most load-bearing concern. The method's formula (Eq. 6–7) has no other learned or adaptive components; the centroids are simple averages of training features, and the threshold is set by the standard TPR-95 convention. Therefore the only thing that separates NCDD from a fixed distance combination is the choice of α1 and α2, and that choice is made using a proxy that is not from the disease classes the method is meant to detect. The paper does not report the chosen hyperparameter values, does not provide sensitivity analysis on real OOD, and gives no error bars, so it is impossible to tell whether the reported improvements over baselines reflect a genuine property of the score or an artifact of tuning on corrupted images. This is an internal-validity issue, not a disagreement with consensus, and it directly conditions the reproducibility and generalizability of the central claim. A real-OOD validation split would settle it. Therefore I agree with the reader's conditional verdict and recommend no change: the paper should be accepted only after the hyperparameter tuning is validated on real OOD data and the values/error bars are reported.","tokens_in":13375,"tokens_out":3793,"duration_ms":35795,"concrete_test":"Hold out one or more real OOD classes (e.g., esophagitis and polyps) from the Kvasirv2 and Gastrovision test sets; tune α1 and α2 on that held-out real-OOD validation split, then evaluate NCDD on the remaining OOD classes across all four backbones. Compare the resulting AUC and FPR95 with the synthetic-tuned values and with the best baseline, and report the selected α1/α2 values along with 95% confidence intervals over at least 5 random seeds. If the real-OOD-tuned weights differ substantially, or if the performance advantage over baselines disappears, the synthetic proxy is the load-bearing assumption and the central claim is weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that NCDD reliably separates GI abnormalities from healthy landmarks across architectures. The score in Eq. (6) depends on α and β, which are defined in Eq. (7) through α1 and α2. The paper states in Section IV.A.2 that these hyperparameters are tuned using synthetic validation OOD data, created by replacing a random rectangular region in ID validation images with random values and adding speckle noise following Hendrycks et al. [40]. This synthetic OOD is therefore corrupted ID imagery, not esophagitis, polyps, ulcerative colitis, or other real disease classes that constitute the test OOD. If the relative magnitudes of nearest versus non-nearest centroid distances for these artificial corruptions differ from those of real diseases, the tuned weights will be miscalibrated for the actual OOD test distribution. No final α1/α2 values are reported for any model or dataset, and Table VII only shows FPR values on synthetic validation OOD for ResNet-18 on Kvasir2. Moreover, the improvement of the full weighted score over simpler distances is modest: FPR95 drops from 33.60 (Dµm alone) to 30.86 (full NCDD) in Table VI. Thus the claimed gains rest on a tuning procedure that has not been validated against any real OOD data, and the method's only adaptive component is selected on a proxy distribution that may not reflect the target distribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper formulates gastrointestinal abnormality detection as an out-of-distribution (OOD) problem, treating healthy anatomical landmarks as in-distribution (ID) and pathologies (esophagitis, polyps, ulcerative colitis, etc.) as OOD. It proposes a post-hoc scoring method, Nearest Centroid Distance Deficit (NCDD), which combines the distance from a test feature vector to its nearest class centroid with the sum of distances to all other centroids, weighted by functions of the feature vector's L1 norm. The method is evaluated on Kvasirv2 and Gastrovision across four backbones (ResNet-18, ViT, DeiT, MLP-Mixer) and compared with MSP, ODIN, Energy, Entropy, MaxLogit, KNN, BLOOD, Neco, and FDBD. The main claim is that NCDD outperforms existing methods on AUC and FPR95 across these settings.","tokens_in":13743,"tokens_out":5127,"duration_ms":40442,"significance":"If the reported results are robust, the paper makes a useful contribution to medical OOD detection by (i) reframing GI abnormality detection as a post-hoc OOD task, (ii) introducing a simple, architecture-agnostic scoring heuristic that captures near-OOD structure, and (iii) releasing the code for reproducibility. The hypothesis that OOD examples are more equidistant from all class centroids, while ID examples are closer to the nearest centroid, is plausible and relevant for medical domains with overlapping features. However, the central claim of uniform superiority is not fully supported by the results, and the only adaptive component of the score is tuned on a synthetic OOD proxy rather than on real disease classes. These issues do not invalidate the approach but require substantial revision in claims and validation.","major_comments":[{"comment":"The abstract and the Conclusion state that NCDD outperforms prior methods on AUC and FPR95 over four model architectures, but Table II for ResNet-18 on Kvasirv2 shows NCDD's AUC (85.68) is lower than MSP (87.57), Entropy (87.55), and KNN (86.59). The claim of consistent superiority is therefore not accurate and should be qualified per architecture and dataset, or replaced with a summary statistic such as mean rank that supports a more moderate statement.","section":"Abstract and Conclusion (Section VI)"},{"comment":"The hyperparameters α1 and α2 in Eq. (7) are tuned on synthetic validation OOD data created by replacing a random rectangle with random values and adding speckle noise following Hendrycks et al. [40]. This proxy consists of corrupted ID imagery, not the real GI abnormality classes (esophagitis, polyps, ulcerative colitis) used as test OOD. The final α1/α2 values are not reported for any model or dataset, and Table VII only shows FPR on synthetic validation OOD for ResNet-18 on Kvasirv2. The authors should report the tuned values and assess the sensitivity of the test results to these choices, or justify that the synthetic proxy is representative of real OOD.","section":"Section IV.A.2"},{"comment":"Table VI shows that the full NCDD score reduces FPR95 from 33.60 (using Dµm alone) to 30.86, a modest absolute improvement. The paper does not report standard deviations or multiple seeds, so it is unclear whether this gain is statistically significant. Since the central claim relies on this improvement, the authors should provide variance estimates or a paired comparison across multiple runs.","section":"Section V.C, Table VI"},{"comment":"The NCDD score in Eq. (6) is introduced as a heuristic without a formal derivation, and the underlying claim that OOD samples have similar nearest and non-nearest centroid distances is supported only by a single ablation for ResNet-18 on Kvasirv2 (Table VI). The paper would be strengthened by verifying this property across all four backbones and both datasets, and by analyzing the score's behavior when the classifier's feature space is not well clustered.","section":"Section III.B"}],"minor_comments":[{"comment":"The abstract uses 'NCCD' once ('we propose a novel nearest-centroid distance deficit (NCCD)') while the rest of the paper and the method name use 'NCDD'; please unify.","section":"Abstract"},{"comment":"The heading 'Quantative Results' contains a typo; it should be 'Quantitative Results'.","section":"Section V.A"},{"comment":"The notation Dµm is defined as the sum of distances to all non-nearest centroids, but the text says 'sum of its distances to all other centroids'; please clarify that it is the sum, not the mean, and make the index set explicit.","section":"Equation (5)"},{"comment":"The references to Hendrycks et al. appear as 'Hendricks et al.' in two places; please correct to 'Hendrycks et al.'.","section":"Section II"},{"comment":"The row for '−Dµn' shows negative values for the mean score; the sign convention and the interpretation of negative scores should be stated explicitly.","section":"Table VI"},{"comment":"The t-SNE plot is difficult to parse in grayscale; consider using distinct markers and a colorblind-friendly palette to distinguish classes.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the empirical study is substantial, but the discrepancy between the abstract/conclusion and Table II, together with the tuning of α1/α2 on synthetic OOD, needs to be addressed before acceptance. If the authors provide qualified claims and additional sensitivity analysis, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nThe paper introduces NCDD, a post-hoc OOD score for gastrointestinal endoscopy. The construction is new: after training a classifier on healthy anatomical landmarks, it computes the Euclidean distance from a test feature to each class centroid, takes the distance to the nearest centroid, sums the distances to all other centroids, and then computes a weighted deficit between those two quantities. The weights are log-scaled by the L1 norm of the feature, with two hyperparameters α1 and α2. The paper evaluates on Kvasirv2 and GastroVision across ResNet-18, ViT, DeiT, and MLP-Mixer, against a solid baseline set.\n\nThe good: the problem is real—GI abnormalities overlap heavily with healthy landmarks, making near-OOD detection hard. The method is simple, post-hoc, and architecture-agnostic, which suits clinical use. The ablation in Table VI supports the intuition that both nearest and non-nearest distances contribute. The authors also provide per-class FPR95 analysis and qualitative examples. The synthetic validation set for hyperparameter tuning is a reasonable stopgap given limited real OOD validation data, and it follows Hendrycks et al.'s outlier exposure framework.\n\nThe soft spots: First, the conclusion overclaims. The abstract and conclusion state NCDD outperforms prior methods on AUC and FPR95 across all architectures, but Table II shows that for ResNet-18 on Kvasirv2, NCDD's AUC (85.68) is below MSP (87.57), Entropy (87.55), and KNN (86.59). That is a factual error in the summary. Second, the hyperparameters α1 and α2 are tuned on synthetic OOD (random rectangles plus speckle), and the paper never reports the values actually used for each dataset/model. Table VII only shows FPR on the synthetic validation set for ResNet-18 on Kvasirv2. Without those values, the method is not fully reproducible, and there is a real risk the tuned weights do not transfer to diseases like esophagitis or polyps. Third, the improvement from the full weighted score over the unweighted deficit is modest—FPR95 goes from 33.60 to 30.86 in Table VI—so the method's advantage is not large. Fourth, no error bars or multiple runs are reported, which is risky for claims of a few percentage points.\n\nOverall, the core idea is plausible and the evaluation is mostly honest, but the paper needs revision: fix the overclaim, report α1/α2 and standard deviations, and ideally validate the weights on a held-out real OOD class. I would send it out for peer review—it is a serious submission with a novel scoring construction and a relevant clinical problem.","headline":"A useful new distance-deficit score for GI OOD detection, but the conclusion overclaims on ResNet-18 Kvasirv2 and the alpha tuning is on synthetic corruptions; worth peer review after revision.","tokens_in":14262,"tokens_out":3915,"would_cite":true,"duration_ms":31190,"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":"This paper claims that a post-hoc score, the nearest-centroid distance deficit, separates healthy gastrointestinal landmarks from abnormalities in endoscopy images better than existing out-of-distribution scores across several…","keywords":["out-of-distribution detection","nearest centroid","gastrointestinal endoscopy","feature space","near-OOD","medical image reliability","deep learning classification"],"falsifier":"Run the same four backbones on Kvasir2 and GastroVision with a validation protocol that tunes alpha1 and alpha2 only on held-out real OOD disease classes rather than on synthetic noise; if NCDD's AUC and FPR95 advantage over KNN and FDBD largely disappears on the held-out class, the paper's central claim is not supported for true near-OOD disease.","tokens_in":13205,"feed_emoji":"🩺","tokens_out":6871,"duration_ms":57387,"temperature":0.7,"pith_summary":"This paper tries to establish that a simple, post-hoc score computed from distances to class centroids in a trained classifier's feature space can flag disease as out-of-distribution in endoscopy images, where abnormalities visually overlap with healthy landmarks. The proposed NCDD score subtracts the distance to the nearest class centroid from a weighted sum of distances to all other centroids, with the weights scaled by the logarithm of the feature vector's L1 norm. Across ResNet-18, ViT, DeiT, and MLP-Mixer backbones on the Kvasir2 and GastroVision benchmarks, the paper reports AUC and FPR95 improvements over standard OOD scores such as MSP, ODIN, Energy, KNN, Neco, and FDBD. The practical stake is that a classifier trained only on normal anatomical landmarks could still alert clinicians to unseen or emerging gastrointestinal abnormalities without retraining or extra annotations.","feed_headline":"New centroid score flags unseen GI disease in endoscopy","feed_subtitle":"A weighted gap between nearest and other class centroids catches abnormalities in endoscopy images without retraining.","key_machinery":"The central object is the nearest-centroid distance deficit. For a test feature vector $z$ from the penultimate layer and class centroids $\\mu_c = \\frac{1}{N_c}\\sum_i z_i^c$, the method computes Euclidean distances $D_c = \\|\\mu_c - z\\|_2$, takes the nearest-centroid distance $D_{\\mu_n} = \\min_c D_c$, and the sum of distances to every other centroid $D_{\\mu_m} = \\sum_{c \\neq \\arg\\min_c D_c} D_c$. The NCDD score is $\\alpha D_{\\mu_m} - \\beta D_{\\mu_n}$, where $\\alpha = \\log(\\|z\\|_1 / 10^{\\alpha_1})$ and $\\beta = \\log(\\|z\\|_1 / 10^{\\alpha_2})$, with $\\alpha_1$ and $\\alpha_2$ tuned on synthetic validation OOD images. This machinery converts the cluster geometry induced by cross-entropy training into a one-dimensional score: in-distribution examples show a large nearest-versus-non-nearest gap, while out-of-distribution examples show a small or negative gap.","core_discovery":"On the paper's own terms, the discovery is a geometric regularity in the feature space of cross-entropy-trained endoscopy classifiers: in-distribution images sit closer to their ground-truth class centroid than to any other centroid, whereas out-of-distribution abnormalities, even near-OOD ones that overlap visually with healthy tissue, sit roughly equally far from all centroids. Exploiting that asymmetry, the NCDD score measures the gap between distance to the nearest centroid and the sum of distances to non-nearest centroids, then thresholds the gap to decide ID versus OOD. The paper further claims that scaling these two terms by logarithms of the feature norm makes the score more sensitive to individual OOD examples, and that this yields higher OOD detection accuracy on Kvasir2 and GastroVision than existing logit-based and feature-based scores across four architectures.","pith_inferences":["A natural extension the paper does not run is per-disease generalization: train on all but one abnormality and test NCDD on the held-out disease, which would directly measure whether the score generalizes to genuinely unseen pathologies rather than the fixed OOD classes in the benchmarks.","Because the score needs only penultimate-layer features and class centroids, it should transfer to other medical imaging domains with a small number of well-clustered in-distribution classes; this is an inference beyond the paper's gastrointestinal experiments.","The log-norm weighting draws on an existing observation that feature norms behave differently for OOD inputs; an untested consequence is that NCDD may be sensitive to input perturbations or preprocessing changes that alter feature magnitudes, which would be worth checking before clinical deployment."],"forward_implications":["A model trained only on normal anatomical landmarks can be repurposed for abnormality flagging without retraining, simply by comparing new images with stored class centroids.","The score works across convolutional, transformer, and MLP-Mixer backbones, suggesting the geometric signal is not tied to one architecture.","Near-OOD cases such as esophagitis overlapping with the Z-line class are caught better than by nearest-neighbor-only feature-space scores.","Low FPR95 means fewer healthy cases are wrongly escalated to clinicians, which supports human-in-the-loop endoscopy workflows."],"supporting_citations":[{"why":"Supplies the Kvasir2 benchmark, with Z-line, cecum, and pylorus as in-distribution classes and the five abnormality classes as out-of-distribution.","marker":"[18]"},{"why":"Supplies the GastroVision benchmark, with 11 normal or anatomical classes as in-distribution and 16 pathological or therapeutic classes as out-of-distribution.","marker":"[38]"},{"why":"Establishes the feature-space observation that ID data cluster near class centroids and serves as a main nearest-neighbor baseline for comparison.","marker":"[30]"},{"why":"Supplies the justification that cross-entropy training compacts intra-class features and separates inter-class features, on which the centroid-distance hypothesis rests.","marker":"[36]"},{"why":"Provides the logarithmic L1-norm weighting idea that NCDD adapts for its alpha and beta weights.","marker":"[37]"},{"why":"Provides the synthetic validation OOD construction, random rectangular corruption plus speckle noise, used to tune the hyperparameters alpha1 and alpha2.","marker":"[40]"}],"fun_headline_variants":["Centroid gap flags unseen GI disease in endoscopy","Nearest-centroid deficit catches OOD in GI images","Distance to class centroid detects out-of-distribution GI","NCDD score spots unseen diseases in endoscopy","Centroid-distance trick improves GI OOD detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's load-bearing premise is that the synthetic validation images made by corrupting normal images with random rectangles and speckle noise resemble real gastrointestinal abnormalities closely enough to tune the two score weights; if real diseases occupy a different part of feature space, the tuned weights could be miscalibrated and the reported gains could shrink.","fun_headline_variants_meta":{"raw":{"variants":["Centroid gap flags unseen GI disease in endoscopy","Nearest-centroid deficit catches OOD in GI images","Distance to class centroid detects out-of-distribution GI","NCDD score spots unseen diseases in endoscopy","Centroid-distance trick improves GI OOD detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000163,"raw_usage":{"total_tokens":1258,"prompt_tokens":977,"completion_tokens":281,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":593,"completion_tokens_details":{"reasoning_tokens":205}},"tokens_in":593,"tokens_out":281,"duration_ms":2973,"temperature":1.0,"reasoning_tokens":205,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:17:26.518483+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same four backbones on Kvasir2 and GastroVision with a validation protocol that tunes alpha1 and alpha2 only on held-out real OOD disease classes rather than on synthetic noise; if NCDD's AUC and FPR95 advantage over KNN and FDBD largely disappears on the held-out class, the paper's central claim is not supported for true near-OOD disease.","supporting_citations":[{"cited_title":"Kvasir: A multi-class image dataset for computer aided gastrointestinal disease detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the Kvasir2 benchmark, with Z-line, cecum, and pylorus as in-distribution classes and the five abnormality classes as out-of-distribution."},{"cited_title":"Gastrovision: A multi-class endoscopy image dataset for computer aided gastrointestinal disease detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the GastroVision benchmark, with 11 normal or anatomical classes as in-distribution and 16 pathological or therapeutic classes as out-of-distribution."},{"cited_title":"Out-of-distribution detection with deep nearest neighbors,","cited_arxiv_id":null,"evidence_quote":"Establishes the feature-space observation that ID data cluster near class centroids and serves as a main nearest-neighbor baseline for comparison."},{"cited_title":"On the separability of classes with the cross- entropy loss function,","cited_arxiv_id":null,"evidence_quote":"Supplies the justification that cross-entropy training compacts intra-class features and separates inter-class features, on which the centroid-distance hypothesis rests."},{"cited_title":"Decoupling maxlogit for out-of-distribution detection,","cited_arxiv_id":null,"evidence_quote":"Provides the logarithmic L1-norm weighting idea that NCDD adapts for its alpha and beta weights."},{"cited_title":"Deep anomaly detec- tion with outlier exposure,","cited_arxiv_id":null,"evidence_quote":"Provides the synthetic validation OOD construction, random rectangular corruption plus speckle noise, used to tune the hyperparameters alpha1 and alpha2."}],"review_version":1}