{"id":"26686841-9640-416b-9e11-0d62726eddff","arxiv_id":"2506.14367","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A hybrid VGG16-DenseNet121 model with explainable heatmaps reached 91.33% accuracy on a small three-class brain MRI dataset.","lead":"This paper combines two standard image-classification networks, VGG16 and DenseNet121, to label brain MRI scans as tumor, normal, or Alzheimer's, reporting 91.33% test accuracy. The authors add Grad-CAM and Integrated Gradients heatmaps to show which image regions drove each prediction.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 91.33% test accuracy rests on a slice-level random split of BraTS volumes; without patient-disjoint splits, within-patient correlation can inflate the result.","rationale":"The reader's weakest_assumption correctly identifies the most load-bearing issue: the stratified random split is performed on slices, not on patients or volumes. Section 3.2 explicitly describes slicing BraTS volumes, and Section 3.3 describes only 'stratified sampling' with no mention of grouping. Because BraTS volumes contain many highly similar axial slices, a slice-level split allows the same patient's data to appear in both training and testing. This makes the 91.33% test accuracy potentially an artifact of within-patient memorization rather than evidence of generalizable multi-class brain disease classification. The paper's central claim — that the hybrid fusion outperforms standard baselines on this dataset — is only meaningful if the evaluation protocol enforces patient independence. The lack of code or data-release artifacts makes the issue impossible to verify from the paper alone, but the concern is concrete and testable. I agree with the reader's CONDITIONAL verdict: the result may be true, but the paper must report patient-level grouping and rerun the evaluation on a patient-disjoint split before the accuracy number can be accepted at face value.","tokens_in":7444,"tokens_out":1962,"duration_ms":22954,"concrete_test":"Ask the authors to report the number of unique BraTS patient/volume IDs in each split and to rerun the same pipeline using GroupShuffleSplit on patient IDs (70/20/10 at the volume level, then extract slices from each volume into its assigned split), keeping architecture and hyperparameters fixed. If the patient-disjoint test accuracy and macro F1 are materially lower (e.g., more than 5 points below 91.33%), the published numbers reflect within-patient leakage rather than generalization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 extracts 2D axial slices from BraTS 2021 volumes, and Section 3.3 then uses stratified random sampling at the slice level without any patient/volume identifier. Under this protocol, slices from the same BraTS volume can appear in both training and test sets. Adjacent T1-weighted axial slices from one volume are highly correlated (same anatomy, same scanner, same intensity distribution), so a model can achieve high test accuracy by memorizing patient-specific patterns rather than learning a generalizable disease classifier. The reported 91.33% accuracy (Table 3) and macro F1 of 90% (Table 4) therefore cannot be interpreted as evidence that DGG-XNet generalizes to new patients. The load-bearing assumption is that the test set is independent of the training set; if slices from the same patient are shared across splits, the central comparison against VGG16 and DenseNet121 is also confounded because all baselines would enjoy the same leakage. This is not an internal inconsistency, but it is a correctness risk for the claimed clinical usefulness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes DGG-XNet, a hybrid deep learning model that fuses VGG16 and DenseNet121 feature maps via global average pooling, for three-class brain MRI classification (Tumour, Normal, Alzheimer's). The model is trained on T1-weighted slices from BraTS 2021 and a Kaggle Alzheimer's dataset, balanced to 500 samples per class and split 70/20/10 by stratified sampling. The authors report 91.33% test accuracy, macro precision/recall/F1 around 90%, and compare against eight CNN baselines. Grad-CAM and Integrated Gradients are used to illustrate model explanations.","tokens_in":7661,"tokens_out":5813,"duration_ms":57034,"significance":"If the evaluation were reliable, the paper would offer a useful empirical data point on hybrid feature fusion for multi-class brain MRI classification, with an interpretability component that is valuable for clinical CAD. The architecture choice is reasonable, and the comparison to several baselines addresses a practical question. However, the current manuscript does not establish patient-level generalization because the split is slice-based, the tumor labeling is not per-slice verified, and the headline accuracy is inconsistent with the reported per-class metrics. These issues are central to the paper's main claim.","major_comments":[{"comment":"The test split is performed at the slice level with stratified random sampling, but the dataset is constructed by extracting multiple 2D axial slices from each BraTS 2021 volume (Section 3.2). Because no patient or volume identifier is used to group slices, slices from the same patient can appear in both training and test sets. Adjacent T1-weighted slices from one volume are highly correlated, so the model can memorize patient-specific intensity patterns; the reported 91.33% accuracy (Table 3) and the comparison to VGG16/DenseNet121 therefore do not establish generalization to new patients. Please re-split at the patient/volume level, state the number of patients in train/validation/test, and re-run all models in Table 3 under this protocol.","section":"Section 3.3 / Table 2"},{"comment":"BraTS 2021 Task 1 volumes are labeled as a single 'Tumour' class for every extracted 2D slice. In a tumor patient's volume, many axial slices (especially at the top and bottom of the brain) contain no tumor tissue; labeling all slices as Tumour is a patient-level label, not a slice-level diagnosis. Since BraTS 2021 provides segmentation masks, please use them to identify slices that actually contain tumor, or explicitly justify patient-level labeling and show that the selected slices are representative. Without this, the Tumour class definition and the measured accuracy are ambiguous.","section":"Section 3.2"},{"comment":"The Alzheimer's dataset is described as having four categories (Non-Demented, Very Mild Demented, Mild Demented, Moderate Demented), but the three-class setup uses only 'Alzheimer's' and 'Normal'. The mapping from the original four classes to these two is never stated, and the class counts after downsampling are given only as 500 per class. Please specify which original labels become 'Alzheimer's', whether 'Non-Demented' serves as 'Normal', and report the per-subclass counts. Otherwise it is unclear whether the Normal and Alzheimer's classes come from the same acquisition protocol and whether 'Normal' in the Alzheimer's dataset is comparable to the healthy tissue seen in BraTS.","section":"Section 3.1 / 3.2"},{"comment":"The reported metrics are internally inconsistent. On a balanced test set of 50 samples per class, accuracy equals the average of the class recalls, i.e., (1.00 + 0.87 + 0.84)/3 = 90.33%, not 91.33%. The abstract states that precision, recall, and F1-score all exceed 91%, but Table 4 gives a macro F1 of 0.90 and per-class scores of 0.82-1.00. Please correct the numbers and provide the confusion matrix with integer counts so that the accuracy, recalls, and precisions can be cross-checked.","section":"Table 4 / Abstract"},{"comment":"The claim that DGG-XNet outperforms VGG16, DenseNet121, and other baselines is not controlled. Table 3 lists only accuracy values; the manuscript does not state whether each baseline was trained with the same optimizer, learning rate, batch size, early-stopping criterion, data balancing, and train/validation/test split. Without identical training protocols, the accuracy differences cannot be attributed to the proposed fusion architecture. Please describe the common training protocol and report per-model hyperparameter settings.","section":"Section 4 / Table 3"}],"minor_comments":[{"comment":"The citation '[6]' for Explainable AI refers to the ImageNet paper; please cite the original Grad-CAM and Integrated Gradients references at the point of use.","section":"Section 3.6"},{"comment":"The ROC curves are shown but no AUC values are reported in the text; list the numerical AUC for each class.","section":"Figure 4b"},{"comment":"The preprocessing description says images are 'resized to 224x224x3 and normalized'; state the normalization scheme (e.g., min-max, z-score) and whether normalization parameters were fitted on the training set only.","section":"Section 3.2"},{"comment":"The integrated-gradients formula would benefit from standard notation for the integration path, and reference [27] is for IG2 rather than the standard integrated-gradients method described; please cite the appropriate original reference.","section":"Equation (11)"},{"comment":"The manuscript does not include a data availability statement or mention whether code will be released; adding these statements would improve reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"I am not convinced the slice-level leakage is intentional, but it is a serious correctness risk for the central claim. The paper will need a full re-run with a patient-disjoint split, corrected and internally consistent metrics, and a controlled baseline comparison before publication. I would also ask the editor to verify the dataset version: the citation 'BraTS 2021 Task 1' is the segmentation benchmark, and using it for classification requires clear per-slice label derivation from the provided segmentation masks."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a standard VGG16+DenseNet121 feature-concatenation application, not a new method. The reported numbers are internally consistent, and the paper is clearly written, but the central accuracy claim depends on an unverified data split: BraTS volumes are sliced into 2D axial images, and then the train/test split is done at the slice level without grouping by patient or volume. Slices from the same patient can appear in both training and testing. Because adjacent slices from one volume are highly correlated, the model can exploit within-patient memorization. The stress-test note is correct, and it is the load-bearing issue here.\n\nWhat the paper does well: it gives a complete pipeline (preprocessing, balanced 500-per-class dataset, stratified split into 1050/300/150), compares against eight baselines, reports macro metrics, and includes Grad-CAM plus Integrated Gradients. The accuracy figures in Table 3 and Table 4 are internally consistent with the confusion matrix. The authors also cite the relevant literature, though the comparison table is a bit thin.\n\nSoft spots, in order of severity. First, the split. Section 3.3 says 'stratified sampling' at the dataset level, but gives no patient or volume identifier. If any BraTS volume contributes slices to both train and test, the test accuracy is inflated, and the comparison against VGG16 and DenseNet121 is also confounded because all baselines use the same leaky split. This is a load-bearing flaw, not a minor one. Second, BraTS slices are labeled Tumour without per-slice verification; some slices near the volume edge may contain little or no tumor, adding label noise. Third, no code or weights are released, no confidence intervals are given, and the test set is only 150 images. Fourth, the novelty is low: concatenating two ImageNet feature extractors with global average pooling is a commonplace transfer-learning pattern.\n\nMy verdict: the paper is not nonsense, but it does not currently support the conclusion that DGG-XNet is a reliable CAD tool. If the split is patient-disjoint, the accuracy might be somewhat lower but still plausible; if not, the result is suspect. This is fixable. A serious referee should see it because the flaw is concrete and the fix is straightforward. I would not cite the current version, but I would bring it to a reading group as a cautionary example of slice-level leakage.\n\nRecommendation: major revision with patient-level splitting, per-slice label verification, and code release, or rejection if the authors cannot redo the evaluation.","headline":"Routine feature-fusion paper whose 91% accuracy rests on a slice-level random split that may leak patient information into the test set.","tokens_in":8213,"tokens_out":2529,"would_cite":false,"duration_ms":26908,"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":"A VGG16 and DenseNet121 feature fusion reaches 91.33% test accuracy on a three-class brain MRI task, outperforming both single backbones.","keywords":["brain MRI classification","hybrid deep learning","VGG16","DenseNet121","global average pooling fusion","Grad-CAM explainability","Alzheimer's disease","brain tumor"],"falsifier":"Re-run the exact training protocol but split the data by patient or 3D volume such that every slice from one volume lands in a single partition; if accuracy falls well below 91.33%, or if a nearest-neighbor check shows train/test slices from the same volume, the reported number reflects identity leakage rather than generalizable classification.","tokens_in":7228,"feed_emoji":"🧠","tokens_out":6720,"duration_ms":60822,"temperature":0.7,"pith_summary":"The paper proposes DGG-XNet, a hybrid deep learning model that runs every MRI slice through two pre-trained convolutional backbones, VGG16 and DenseNet121, pools the resulting feature maps with global average pooling, concatenates the pooled vectors, and sends them through dense layers to a three-way softmax (Tumour, Normal, Alzheimer's). Trained on a balanced dataset built from the BraTS 2021 tumor volumes and a public Alzheimer's MRI collection, the model reports 91.33% test accuracy, with macro-averaged precision, recall, and F1-score near 0.90, on a 150-image balanced test set. The paper's claim is that this feature-level fusion beats every single backbone it compares against, including VGG16 at 84.67% and DenseNet121 at 82.67%, and that Grad-CAM plus Integrated Gradients show the decisions rest on interpretable brain regions. A sympathetic reader would take this as evidence that fusing complementary CNN feature extractors is a promising, more transparent route for computer-aided screening of neurological disease.","feed_headline":"Fusing VGG16 and DenseNet121 hits 91.33% on brain MRI","feed_subtitle":"The fused model beats single VGG16 and DenseNet121 baselines on tumor, normal, and Alzheimer's MRI scans.","key_machinery":"The load-bearing mechanism is dual-branch feature fusion with global average pooling. Each input image $x$ is passed through VGG16 and DenseNet121; the pooled branch outputs are concatenated into $F = \\mathrm{Concat}(\\mathrm{GAP}(VGG16(x)), \\mathrm{GAP}(DenseNet121(x)))$, and $F$ is transformed by batch-normalized dense layers with ReLU, dropout, and a softmax head. The concat fusion is what allows the final classifier to combine VGG16's spatial hierarchy with DenseNet121's densely reused features before making the three-way decision.","core_discovery":"The central discovery is that fusing VGG16 and DenseNet121 at the feature-vector level improves three-class brain MRI classification beyond either backbone alone: DGG-XNet reaches 91.33% test accuracy while VGG16 alone reaches 84.67% and DenseNet121 alone reaches 82.67% on the same balanced 150-image test set. The authors attribute the gain to VGG16's strong hierarchical spatial features complementing DenseNet121's dense connectivity and feature reuse, together with ImageNet-initialized transfer learning, downsampling to 500 images per class, and early stopping.","pith_inferences":["Because the split described in the paper does not group slices by patient or 3D volume, the 91.33% figure could be inflated by same-patient slices appearing in both training and testing; a patient-level evaluation would settle this directly.","The fusion advantage could be probed by ablating the concatenation against attention-weighted fusion or single-branch global average pooling, to see whether the gain is the fusion itself or simply added model capacity.","Grad-CAM and Integrated Gradients maps are shown qualitatively; their faithfulness could be quantified with insertion/deletion or pointing-game metrics.","Because BraTS volumes and the Alzheimer's 2D set differ in acquisition and content, testing DGG-XNet on a third unseen dataset, or with per-scanner cross-validation folds, would show whether the 91.33% transfers across imaging protocols."],"forward_implications":["If the reported accuracy holds, feature-fusion hybrids of this kind are a stronger starting point than single CNNs for multi-class neurological screening from 2D MRI slices.","The same training recipe, including balanced downsampling, ImageNet initialization, and early stopping, yields a model whose Grad-CAM and Integrated Gradients maps point to localized brain structures.","The comparison table implies that a family of fusion architectures can be swept by swapping backbones while keeping the global-average-pooling, concatenation, and dense-head design fixed.","The class-wise metrics show the remaining error concentrates on the Alzheimer's class, with precision 0.82 and recall 0.84, while Tumour is classified perfectly at 1.00.","The paper's stated future directions, including 3D volumetric analysis, added modalities, data augmentation, and domain adaptation, are natural next tests of whether the fusion benefit transfers beyond the current 2D slice setting."],"supporting_citations":[{"why":"Supplies the DenseNet121 backbone and the dense-connectivity rationale for feature reuse and gradient flow.","marker":"[10]"},{"why":"Supplies the VGG16 backbone and the hierarchical spatial feature representations the fusion relies on.","marker":"[23]"},{"why":"Provides the ImageNet pretrained weights that initialize both backbones through transfer learning.","marker":"[6]"},{"why":"Introduces global average pooling, the operation that compresses each branch's feature maps before concatenation.","marker":"[15]"},{"why":"Defines Grad-CAM, the gradient-based localization method used to visualize the regions driving the predictions.","marker":"[21]"},{"why":"Provides the BraTS 2021 brain-tumor MRI volumes from which the 2D tumor slices are extracted.","marker":"[3]"}],"fun_headline_variants":["Hybrid VGG16-DenseNet121 hits 91.33% on brain MRI","Fused VGG16+DenseNet121 beats individual nets on brain MRI","Explainable AI fusion model for brain disease hits 91.33%","DGG-XNet fuses VGG16 & DenseNet121 for brain MRI classification","91.33% accuracy: hybrid VGG16-DenseNet121 on brain MRI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the stratified random split of individual 2D MRI slices into training and test sets keeps the test set independent; since slices are drawn from the same 3D volumes without patient-level grouping, slices from the same patient can appear on both sides, which would let the model memorize patients rather than learn the disease.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid VGG16-DenseNet121 hits 91.33% on brain MRI","Fused VGG16+DenseNet121 beats individual nets on brain MRI","Explainable AI fusion model for brain disease hits 91.33%","DGG-XNet fuses VGG16 & DenseNet121 for brain MRI classification","91.33% accuracy: hybrid VGG16-DenseNet121 on brain MRI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000775,"raw_usage":{"total_tokens":3382,"prompt_tokens":853,"completion_tokens":2529,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":2421}},"tokens_in":469,"tokens_out":2529,"duration_ms":18486,"temperature":1.0,"reasoning_tokens":2421,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:17:43.707107+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the exact training protocol but split the data by patient or 3D volume such that every slice from one volume lands in a single partition; if accuracy falls well below 91.33%, or if a nearest-neighbor check shows train/test slices from the same volume, the reported number reflects identity leakage rather than generalizable classification.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the DenseNet121 backbone and the dense-connectivity rationale for feature reuse and gradient flow."},{"cited_title":"In: Proceedings of the International Conference on Learning Representations (ICLR)","cited_arxiv_id":null,"evidence_quote":"Supplies the VGG16 backbone and the hierarchical spatial feature representations the fusion relies on."},{"cited_title":"ArXiv (2013)","cited_arxiv_id":null,"evidence_quote":"Introduces global average pooling, the operation that compresses each branch's feature maps before concatenation."},{"cited_title":"In: 2017 IEEE International Conference on Computer Vision (ICCV)","cited_arxiv_id":null,"evidence_quote":"Defines Grad-CAM, the gradient-based localization method used to visualize the regions driving the predictions."}],"review_version":1}