{"id":"1579fa41-4651-482d-bff7-4d93944b883f","arxiv_id":"2507.22274","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"HOG-CNN fuses a 26,244-dimensional HOG descriptor with a frozen CNN embedding and reports accuracy up to 98.5% on APTOS, 92.8% on IC-AMD, and 83.9% on ORIGA.","lead":"HOG-CNN combines handcrafted gradient features with pretrained CNN features to classify retinal diseases, reporting high accuracy on three public datasets. The reported results may be affected by how the best model configuration was selected per dataset.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline numbers come from a different best-performing backbone per dataset, so the claim that a single HOG-CNN model achieves consistently state-of-the-art performance is not established.","rationale":"The reader's weakest assumption correctly identifies the per-dataset backbone variation as the key vulnerability. The paper describes HOG-CNN as a single model, but the appendix reveals that each reported number uses a different pretrained backbone. Without an explicit validation-based selection procedure, the reported results are consistent with best-of-ten selection on the test set, which inflates apparent performance and prevents the reader from assessing generalizability. I also note an internal contradiction supporting this concern: Table .7 shows plain VGG19 reaching 95.65 AUC on 5-class DR, above the reported HOG-CNN 94.2 AUC, while the main text claims the highest AUC. This makes the reported performance depend on both backbone choice and metric choice. A fixed-backbone rerun would settle whether the HOG augmentation itself is responsible for the gains. I am not recommending rejection: the code is public, the datasets are standard, and the hybrid idea is plausible, but the paper should be revised to report a single architecture or an explicit, validation-only model selection protocol. The existing CONDITIONAL verdict remains appropriate.","tokens_in":16370,"tokens_out":7509,"duration_ms":87453,"concrete_test":"Run the public repository with a single fixed backbone (e.g., EfficientNetB2) for all four tasks, using the same splits and hyperparameters, and without per-dataset backbone edits. Report base-CNN and HOG-CNN metrics side by side with bootstrap confidence intervals. If the fixed-backbone HOG-CNN does not beat the best baseline in each task, or if HOG features do not improve over the same backbone alone, then the state-of-the-art claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that HOG-CNN, as one architecture, delivers state-of-the-art or near-state-of-the-art results on all three tasks. The appendix (Tables .6-.9) shows that the headline result for each task is reported with a different pretrained backbone: EfficientNetB3 for binary APTOS, EfficientNetB2 for 5-class APTOS and IC-AMD, and EfficientNetB0 for ORIGA. The paper does not state that the backbone was chosen by a validation-based selection rule. It is therefore consistent with selecting, after observing test performance, the best of the ten configurations listed in the appendix. The effect is not purely cosmetic: in Table .7, plain VGG19 achieves a 5-class DR AUC of 95.65, which is higher than the reported HOG-CNN AUC of 94.2, and adding HOG to VGG19 lowers AUC to 92.26. The main text nevertheless presents 94.2 AUC as the highest. This pattern means the headline performance is a selected maximum over backbones and metrics rather than the behavior of a single model. The claim of consistent high performance is thus load-bearing on an untested assumption about model selection.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes HOG-CNN, a hybrid architecture that concatenates Histogram of Oriented Gradients (HOG) descriptors with features from a frozen pretrained CNN backbone, and evaluates it on three retinal fundus benchmarks: binary and five-class diabetic retinopathy on APTOS 2019, age-related macular degeneration on IChallenge-AMD, and glaucoma on ORIGA. The main text reports state-of-the-art or near-state-of-the-art accuracy, precision, recall, and AUC numbers for each task, and an appendix tabulates results for ten different CNN backbones with and without HOG fusion. The paper argues that the fusion of handcrafted gradient features with deep features yields consistently high performance while remaining lightweight and interpretable.","tokens_in":16599,"tokens_out":5108,"duration_ms":53205,"significance":"If the reported results were produced by a single, well-specified model under controlled comparison, the paper would provide a useful data point on combining handcrafted HOG features with transfer-learned CNN features for retinal screening, supported by public code and a broad set of per-backbone experiments. The appendix's systematic reporting of ten backbones is a strength. However, the central claim of a single HOG-CNN architecture achieving consistent state-of-the-art performance is not currently supported, because the headline numbers are drawn from different backbones across datasets and the comparisons to prior work are made under incompatible evaluation protocols. The empirical contribution is therefore real but overstated, and the paper in its present form does not support the generalization claims in the abstract and conclusion.","major_comments":[{"comment":"The headline result for each dataset is not produced by a single HOG-CNN configuration. The binary APTOS result in Table 3 corresponds to HOG+EfficientNetB3 in Table .6; the five-class APTOS and IC-AMD results correspond to HOG+EfficientNetB2 in Tables .7 and .8; and the ORIGA result corresponds to HOG+EfficientNetB0 in Table .9. The manuscript does not specify a validation-based rule for selecting these backbones, so the reported numbers are consistent with selecting the best of the ten configurations after observing test performance. This is not a cosmetic issue: in Table .7, plain VGG19 achieves a five-class DR AUC of 95.65, above the reported HOG-CNN AUC of 94.2, and in Table .8, HOG+EfficientNetB3 has an IC-AMD AUC of 93.61 while HOG+EfficientNetB2 has 94.49. The text's claim that a single 'HOG-CNN' model performs consistently at the state of the art is therefore not established by the present experiments.","section":"Section 5 and Appendix Tables .6-.9"},{"comment":"The comparisons against prior work use different train/test splits or cross-validation schemes across methods: HOG-CNN uses 80:20 for APTOS and ORIGA while many baselines use 85:15 or 10-fold CV; Table 4 lists no split for MobileNetV2-SVM; Table 5 mixes 70:30, 10-fold, and pretrained setups; and the CycleGAN row in Table 2 uses a different class balance (933:267). No confidence intervals or repeated-run variability are reported. Under these conditions, differences of a fraction of a percentage point, which the text interprets as 'surpassing' (e.g., 98.5 vs 98.4 accuracy in Table 3), cannot be distinguished from split-induced variation. The superiority claims in Section 5 require evaluation on identical splits with confidence intervals or significance tests.","section":"Tables 2-5"},{"comment":"The sentence 'the proposed model achieved an accuracy, precision, and recall of 92.8%, along with an AUC of 94.5%' is contradicted by Table 2, which lists precision 94.8 and recall 72.0 for HOG-CNN on the IC-AMD dataset. This factual error also undermines the subsequent argument that HOG-CNN offers 'a more balanced performance across all metrics' than S2C2L, since the reported recall is ten points below S2C2L's 82.4%.","section":"Section 6, first paragraph"},{"comment":"In the five-class DR setting, the paper reports precision 86.0, recall 78.6, accuracy 83.2, and AUC 94.2 as the HOG-CNN result. Appendix Table .7 shows that no single backbone simultaneously attains these values: HOG+EfficientNetB0 reaches 83.49 accuracy, while HOG+EfficientNetB2 reaches 94.15 AUC, and the two configurations differ in precision and recall. Reporting a composite of two different configurations as one model's result reinforces the selection-bias concern raised above and should be clarified or corrected.","section":"Section 5, DR Detection Results"}],"minor_comments":[{"comment":"The dataset is referred to inconsistently as 'IC-Dataset', 'IChallenge-AMD', and 'IC-AMD'; one name should be used throughout.","section":"Highlights and Section 4.1"},{"comment":"The text describing the HOG pathway reads 'the first dense layer contains 800 ReLU-activated units, the second dense layer contains 256 ReLU-activated units, and the second contains 128 ReLU-activated units'; the last occurrence should be 'the third'.","section":"Section 3.2"},{"comment":"The phrase 'summarized in here 1' should be 'summarized in Algorithm 1'.","section":"Algorithm 1 reference"},{"comment":"In the InceptionResNetV2 row, the precision value '2152' appears to be a typo for 21.52.","section":"Table .8"},{"comment":"The CNN-SVM row lists a class balance of 660:453 for ORIGA, although the dataset is described as containing 650 images; this discrepancy needs clarification.","section":"Table 5"},{"comment":"AUC values are reported with percent signs (e.g., 99.2%), which is nonstandard for a unitless quantity; either report AUC as a decimal or state the convention explicitly.","section":"General reporting"},{"comment":"The related-work section contains an extended discussion of topological data analysis that is not connected to the proposed method or experiments, making the section harder to follow.","section":"Section 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a single-author empirical study with a useful public code link and an informative per-backbone appendix. The main text, however, overstates the results relative to what the appendix actually shows, and the Discussion contains a factual error about the paper's own Table 2. For this journal, I would want the authors to either fix the backbone-selection issue by committing to a single architecture chosen by a pre-registered validation rule, or reframe the paper as an exploratory per-backbone comparison. The cross-dataset comparison tables also need to be put on a common evaluation protocol before the stated state-of-the-art claims can be taken seriously."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent but unremarkable benchmark paper. The architecture is a standard concatenation of frozen CNN features and HOG embeddings; the novelty is limited to applying it to three retinal datasets. The appendix is actually the most informative part—10 backbones with and without HOG on each dataset, plus a code link. That deserves credit.\n\nThe problem is the main text's central claim. The reported \"HOG-CNN\" numbers are not one model. Binary APTOS uses EfficientNetB3; 5-class APTOS and IC-AMD use EfficientNetB2; ORIGA uses EfficientNetB0. The paper never states a validation-based rule for choosing these backbones, so the headline numbers look like the best of ten configurations selected after seeing test results. That is selection bias, not a single architecture's consistent performance. The appendix makes it worse: on 5-class APTOS, plain VGG19 gets 95.65 AUC while HOG+VGG19 drops to 92.26, and the main text reports 94.2 from a different backbone as \"highest.\" So the claim that adding HOG gives complementary strength is not consistently supported.\n\nThere are smaller issues: comparisons against baselines on different splits (80:20 vs 85:15 vs 10-fold), no confidence intervals, and a factual error in the Discussion—it claims recall of 92.8% for IC-AMD, but Table 2 shows 72.0%. The \"interpretable\" claim is never backed by any analysis.\n\nNone of this is fatal to the idea of using HOG as a cheap feature channel. A revised version that pre-specifies one backbone per task or uses nested validation, reports confidence intervals, and compares on identical splits could be a useful lightweight baseline for retinal screening. As is, the performance claims should not be taken at face value.\n\nWho is this for? Someone building a simple baseline in retinal image classification, not someone looking for methodological novelty. I would not cite it in current form. But the appendix has real information, and the core issue is fixable, so a serious referee could work with it. Send it to review, with the expectation of major revision.","headline":"Useful appendix, but the headline numbers are a per-dataset best-backbone selection, so the claim that a single HOG-CNN model performs consistently well does not hold as written.","tokens_in":17115,"tokens_out":4890,"would_cite":false,"duration_ms":58849,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A hybrid HOG-CNN model fuses handcrafted gradient histograms with frozen pretrained CNN features, and the paper claims this raises retinal disease classification accuracy and AUC over pure deep-learning baselines across three public…","keywords":["Retinal disease diagnosis","Histogram of Oriented Gradients","Convolutional neural networks","Feature fusion","Fundus images","Diabetic retinopathy","Glaucoma","Age-related macular degeneration"],"falsifier":"Run the same four tasks with a single pre-registered backbone and without inspecting test labels; if the HOG+CNN variant does not beat the CNN-only variant in accuracy or AUC on held-out folds, the claimed complementarity is refuted.","tokens_in":16147,"feed_emoji":"👁","tokens_out":5190,"duration_ms":52395,"temperature":0.7,"pith_summary":"The paper proposes a hybrid classifier, HOG-CNN, that fuses handcrafted Histogram of Oriented Gradients features with features from a frozen pretrained CNN to classify retinal fundus images. It claims this fusion captures both local texture and edge structure and high-level semantic content, improving discrimination for diabetic retinopathy, age-related macular degeneration, and glaucoma over pure deep learning baselines. On APTOS 2019 the model reaches 98.5% accuracy and 99.2 AUC for binary DR and 94.2 AUC for five-class DR; on IC-AMD it reaches 92.8% accuracy and 94.5 AUC; on ORIGA it reaches 83.9% accuracy and 87.2 AUC. The paper argues that because the HOG pathway needs no data augmentation, segmentation, or domain-specific pretraining, the method is a lightweight, interpretable option for resource-constrained screening.","feed_headline":"HOG+CNN fusion tops 98.5% accuracy on diabetic retinopathy","feed_subtitle":"Handcrafted gradient features plus deep CNN features beat pure deep models on DR, AMD, and glaucoma screening.","key_machinery":"The central mechanism is a two-branch architecture. A grayscale version of each fundus image is converted into a 26,244-dimensional HOG descriptor with 9 orientation bins, 8x8 cells, and 2x2 normalized blocks, then compressed by a three-layer MLP of sizes 800, 256, and 128; in parallel, the RGB image passes through a frozen pretrained CNN with its classification head removed, followed by a 64-filter 3x3 convolution, 2x2 max pooling, flattening, and a 64-unit dense layer. The two embeddings are concatenated into a 192-dimensional vector and fed to two further dense layers with dropout 0.2 before a sigmoid or softmax output. This design lets the model combine local gradient histograms with deep semantic features.","core_discovery":"The central claim is that adding a handcrafted gradient-histogram branch to a deep CNN improves retinal disease classification beyond either feature type alone. In the paper's experiments the fused model beats the CNN-only version on nearly every backbone and dataset combination reported in the appendix, and the headline configuration outperforms published baselines: 98.5% accuracy and 99.2 AUC for binary diabetic retinopathy, 83.2% accuracy with 94.2 AUC for five-class grading, 92.8% accuracy and 94.5 AUC for AMD, and 83.9% accuracy and 87.2 AUC for glaucoma. The author treats this as evidence that local texture information captured by HOG is complementary to the semantic features learned by CNNs in ophthalmic imaging.","pith_inferences":["The appendix tests ten backbones with and without HOG, and the headline numbers are the best per dataset; the one-model framing is strongest if each backbone choice was fixed before test performance was inspected.","A testable extension is to check whether the HOG gain persists on external or out-of-distribution fundus datasets where illumination and resolution vary; if it does, the method's value is robustness rather than only benchmark accuracy.","The reported precision and recall asymmetry on IC-AMD, 94.8% precision versus 72.0% recall, suggests the model is conservative in flagging AMD, so future work could tune the decision threshold to meet clinical recall requirements.","The same dual-branch fusion could be combined with other handcrafted descriptors, such as color histograms or topological features, which the paper leaves to future work and which might improve recall on the minority AMD and glaucoma classes."],"forward_implications":["If the reported results hold under fair comparison, HOG-CNN shows that handcrafted descriptors still add value on top of modern CNN features in medical imaging.","The method reaches near-state-of-the-art accuracy without data augmentation, segmentation, or fine-tuning of the backbone, which shortens training and lowers compute requirements.","Because the HOG branch is computed in grayscale and the CNN branch is frozen, the model can be trained on CPU and deployed in low-resource clinics.","The consistent gains of HOG+CNN over CNN-only in the appendix support the general claim that gradient and texture cues are complementary to deep features across three diseases.","The four-task results establish a single fusion recipe that transfers across DR, AMD, and glaucoma without task-specific engineering."],"supporting_citations":[{"why":"Supplies the Histogram of Oriented Gradients descriptor method that the HOG branch is built on.","marker":"[8]"},{"why":"Provides the APTOS 2019 benchmark dataset used for binary and five-class diabetic retinopathy classification.","marker":"[36]"},{"why":"Provides the ORIGA benchmark dataset used for glaucoma detection.","marker":"[35]"},{"why":"Provides the IChallenge-AMD dataset used for age-related macular degeneration classification.","marker":"[33]"},{"why":"A supervised contrastive learning baseline whose binary DR accuracy and AUC the paper claims to surpass.","marker":"[38]"},{"why":"A self-supervised baseline on IC-AMD with the highest reported AUC, providing the main comparison point the paper must address.","marker":"[55]"},{"why":"A segmentation-based glaucoma model with top ORIGA accuracy; the paper compares its simpler HOG-CNN results against it.","marker":"[42]"},{"why":"An optic-disc localization and glaucoma classification baseline that defines the top ORIGA performance the paper positions against.","marker":"[43]"}],"fun_headline_variants":["Hybrid HOG-CNN model beats pure CNNs on retinal disease","Lightweight HOG-CNN fusion screens DR, glaucoma, AMD","Gradient texture + deep features: efficient retinal disease detector","Fusing HOG with CNN improves retinal disease classification","HOG+CNN fusion hits 98.5% DR accuracy, 99.2 AUC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline results use a different pretrained backbone for each dataset, and the paper never states that those backbone choices were made before seeing test performance, so the claimed consistent advantage could be an artifact of picking the best of several configurations rather than a property of the HOG+CNN fusion.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid HOG-CNN model beats pure CNNs on retinal disease","Lightweight HOG-CNN fusion screens DR, glaucoma, AMD","Gradient texture + deep features: efficient retinal disease detector","Fusing HOG with CNN improves retinal disease classification","HOG+CNN fusion hits 98.5% DR accuracy, 99.2 AUC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000977,"raw_usage":{"total_tokens":4198,"prompt_tokens":1043,"completion_tokens":3155,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":659,"completion_tokens_details":{"reasoning_tokens":3062}},"tokens_in":659,"tokens_out":3155,"duration_ms":27604,"temperature":1.0,"reasoning_tokens":3062,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T11:52:19.603490+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same four tasks with a single pre-registered backbone and without inspecting test labels; if the HOG+CNN variant does not beat the CNN-only variant in accuracy or AUC on held-out folds, the claimed complementarity is refuted.","supporting_citations":[{"cited_title":"Dalal, B","cited_arxiv_id":null,"evidence_quote":"Supplies the Histogram of Oriented Gradients descriptor method that the HOG branch is built on."},{"cited_title":"https://www.kaggle.com/c/ aptos2019-blindness-detection","cited_arxiv_id":null,"evidence_quote":"Provides the APTOS 2019 benchmark dataset used for binary and five-class diabetic retinopathy classification."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the ORIGA benchmark dataset used for glaucoma detection."},{"cited_title":"Fang, et al., ADAM challenge: Detecting age-related mac- ular degeneration from fundus images, IEEE Transactions on Medical Imaging (2022)","cited_arxiv_id":null,"evidence_quote":"Provides the IChallenge-AMD dataset used for age-related macular degeneration classification."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A supervised contrastive learning baseline whose binary DR accuracy and AUC the paper claims to surpass."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A self-supervised baseline on IC-AMD with the highest reported AUC, providing the main comparison point the paper must address."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A segmentation-based glaucoma model with top ORIGA accuracy; the paper compares its simpler HOG-CNN results against it."},{"cited_title":"Latif, S","cited_arxiv_id":null,"evidence_quote":"An optic-disc localization and glaucoma classification baseline that defines the top ORIGA performance the paper positions against."}],"review_version":1}