{"id":"783fef80-7519-4702-902d-7ecd587cef76","arxiv_id":"2412.03084","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A hybrid CNN plus deep classifier reports 100% accuracy on TCGA-LIHC liver cancer grading and 96.71% on a hospital dataset, but the evaluation likely suffers from patch-level data leakage.","lead":"This paper combines a pre-trained image recognition network with a custom deep classifier to automatically grade liver cancer from stained tissue images, reporting near-perfect accuracy on a public dataset and high accuracy on an Indian hospital dataset. The study is a standard transfer-learning benchmark whose numbers would be more credible under strict patient-level splitting.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Patch-level data leakage is the load-bearing weakness: Table I and Section II.C split patches, not slides, so the same TCGA/KMC WSI can contribute to both train and test; the 100% and 96.71% results therefore do not establish slide-level generalization. A slide-grouped split is required.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing issue: statistical independence of the train and test sets is assumed but not established, because the split is at the patch level rather than the slide or patient level. This is not a minor implementation detail; it directly controls whether the reported accuracies reflect generalization to new histopathology slides or memorization of slide-specific staining patterns. The perfect 100%/AUC 1.00 TCGA result and the high KMC result are exactly what one would expect under such leakage. I also note the secondary problem of selecting the best feature extractor per dataset on the test set, which further inflates the chosen model's numbers, but the patch-level split is the more fundamental flaw. The paper's own limitations section does not acknowledge this, and no code is provided that would allow verification of the split. Therefore the rejection is appropriate as submitted; a revision with slide-level grouped splitting and a fixed validation protocol could be reconsidered.","tokens_in":19708,"tokens_out":4579,"duration_ms":50596,"concrete_test":"Rerun the identical pipeline with a group-based split on TCGA-LIHC and KMC, with groups defined by whole-slide image ID (and patient ID where possible) instead of a random patch split, keeping all hyperparameters and augmentation unchanged; report accuracy, sensitivity, specificity, F1, and AUC for hybrid ResNet50 on TCGA and hybrid EfficientNetb3 on KMC. Also report the fraction of test patches whose source slide appears in the training set. If the 100% and 96.71% figures drop materially, the published results are artifacts of patch-level leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a hybrid CNN/ANN architecture with selective fine-tuning outperforms frozen-backbone base models, supported by the headline TCGA (100%) and KMC (96.71%) accuracies. For that claim to be meaningful, the test set must estimate performance on unseen slides. The protocol in Section II.C and Table I splits 3920 TCGA patches and 3005 KMC patches into train/test randomly at the patch level; no grouping by whole-slide image or patient is described anywhere. Patches from the same WSI share staining, scanner, and tissue context, and the static augmentation (flips) applied before the split can place near-duplicates of the same original patch in both partitions. A model can then memorize slide-specific appearance rather than grade-relevant histology, which would directly explain the perfect TCGA result and inflate the KMC result. The paper's stated limitations address parameter count only and never mention this split hazard. Thus the strongest empirical support for the hybrid improvement is not trustworthy as evidence of generalization to new slides.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes a hybrid CNN-ANN architecture for hepatocellular carcinoma (HCC) grade classification from H&E-stained liver histopathology images. The method uses a pre-trained CNN (one of eight backbones) as a feature extractor, freezes its bottom layers, fine-tunes its top layers, and replaces the final classifier with a deep stack of fully connected layers. The authors evaluate the approach on the TCGA-LIHC dataset (3920 patches, three classes), a proprietary KMC dataset (3005 patches, four classes), and the LC25000 colon dataset (10,000 patches, two classes), using five-fold cross-validation and a 90/10 train/test split. They report that the hybrid model with ResNet50 achieves 100% accuracy on TCGA, that the hybrid model with EfficientNetb3 achieves 96.71% accuracy on KMC, and that hybrid models improve accuracy by roughly 2-5 percentage points over the corresponding base models.","tokens_in":19977,"tokens_out":7404,"duration_ms":62117,"significance":"If the reported results were obtained under a statistically valid evaluation protocol, the paper would offer a simple and reproducible recipe—selective fine-tuning of top convolutional layers plus a deep classifier—that improves several well-known backbones for histopathology patch classification. The systematic comparison across eight architectures and three datasets is a useful reference, and the public TCGA and LC25000 results would facilitate replication. However, the significance is currently undercut by severe methodological concerns about data leakage and test-set-based model selection, which make the absolute accuracy numbers and the claimed improvements unreliable as evidence of generalization to unseen slides.","major_comments":[{"comment":"The train/test split is performed at the patch level rather than at the level of whole-slide images or patients. Section II.A describes tiling 491 TCGA WSIs into 1024×1024 patches, and Section III.A and Table I then divide the resulting 3920 patches into 3528 training and 392 test patches randomly. Because patches from the same WSI share staining, scanner, and tissue context, this protocol does not yield an independent test set, and the 100% TCGA and 96.71% KMC accuracies are likely inflated by the model memorizing slide-specific appearance. The authors must re-run the evaluation with a split that assigns all patches from a given WSI (or patient) to the same fold, and report metrics under that protocol.","section":"Section II.A, Section III.A, Table I"},{"comment":"The data augmentation described in Section II.A (random vertical and horizontal flips) is applied before the train/test split of Table I. Since the augmented counts (1220, 1340, 1360) are derived from the original extracted patches (813, 893, 680), flipped versions of the same original patch can appear in both training and test partitions, putting near-duplicates in the test set. The authors should split the data before any augmentation, or ensure that augmented copies of a given original patch are confined to the same partition.","section":"Section II.A and Section III.A"},{"comment":"The best feature extractor is chosen by evaluating all eight candidate models on the same test set and then reporting the top performer (ResNet50 on TCGA, EfficientNetb3 on KMC) as the proposed result. This is a form of test-set fitting: the reported accuracy is the maximum over eight random variables and is therefore optimistically biased. The authors should select the architecture on a separate validation set (or via nested cross-validation) and report performance on a test set held out from all selection decisions.","section":"Section III.C, Tables IV and VI"},{"comment":"The evaluation protocol is unclear: Section II.C describes five-fold cross-validation on the training data, but Section III.A and Table I introduce a separate 90/10 test split, and the results in Tables III-VIII are said to be 'averaged over five-folds.' It is not specified whether the test set is reused across folds or whether the reported numbers are cross-validation fold averages. This ambiguity makes the metrics difficult to interpret and must be resolved by a precise description of the training/validation/test workflow.","section":"Section II.C and Section III.C"},{"comment":"The Limitations paragraph addresses only the high parameter count of the models. It does not mention the absence of slide-level splitting, the risk of augmentation leakage across the train/test boundary, or the optimistic bias from choosing the feature extractor on the test set. These are the most serious threats to the validity of the reported results and should be acknowledged and discussed.","section":"Section IV"}],"minor_comments":[{"comment":"The claim of 'improvement in accuracy of 2% and 4%' is inconsistent with the detailed numbers (1.74 percentage points for TCGA in Section III.C.1 and 4.65 percentage points for KMC in Section III.C.2). Please use percentage points consistently.","section":"Abstract and Section III.C"},{"comment":"The augmentation results in different multipliers per class (1220/813 ≈ 1.5, 1340/893 ≈ 1.5, 1360/680 = 2.0). The exact augmentation strategy should be described, particularly why the type-2 class received a larger number of augmented patches.","section":"Section II.A"},{"comment":"The 'offline system' and 'online system' are introduced without definitions; a one-sentence explanation would help the reader.","section":"Figure 5"},{"comment":"The sentence 'Cosine annealing warm restart [38] learning rate scheduler is used to select the learning rate over all epochs' should read '...to schedule the learning rate...'.","section":"Section III.A"},{"comment":"The colon dataset is labeled 'LC25000' in Table IX but called 'COLON' elsewhere; unify the name.","section":"Table IX"},{"comment":"The statement that the training curves indicate 'the absence of both bias and variance' is not established by the presented plots; consider replacing it with a more cautious description.","section":"Sections III.C.1 and III.C.2"}],"recommendation":"major_revision","confidential_remarks":"The reader's report recommends rejection, and I agree that the current experimental evidence is not trustworthy. However, the core idea—selective fine-tuning with a deep classifier—is not inherently flawed, and the issues are addressable by re-running the experiments with a slide-level split and an unbiased model-selection protocol. I would therefore be willing to see a substantially revised version, provided the authors can either provide such results or clearly restate the claims as patch-level memorization experiments. If the authors cannot perform a slide-level split (e.g., because slide IDs are unavailable for KMC), the paper's claims should be scaled back accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper is a straightforward transfer-learning benchmark with a serious methodological flaw: the train/test split is at the patch level, not the slide or patient level, and the best feature extractor is chosen on the test set. That means the headline 100% TCGA and 96.71% KMC numbers are not evidence of slide-level generalization.\n\nWhat's actually new: not much. The hybrid recipe—pre-trained CNN with selective fine-tuning of top layers plus a deeper ANN classifier—is standard transfer learning, anticipated by the paper's own citations (Talo et al., Kandel and Castelli). The contribution is a parameter scan of eight backbones on three datasets. To the paper's credit, it uses a real external clinical dataset (KMC) for validation, describes preprocessing in detail, and reports error bars, ROC curves, and confusion matrices. That is a solid empirical framework, and the comparison to LiverNet in Table IX is useful context.\n\nThe problem is the split. Table I and Section II.C show that the TCGA dataset is expanded by static flips (3920 patches from 2386 originals) before the 90/10 train/test division. No slide or patient grouping is mentioned. Patches from the same WSI share staining and scanner artifacts, and flipped versions of the same patch can straddle the split. The model can memorize slide-specific appearance, which directly explains the perfect TCGA result. The same hazard applies to the KMC data, which is split without reference to slides. Moreover, the paper selects the best of eight feature extractors per dataset based on the test-set performance, which is fitting to the test set and inflates the reported numbers. The limitations section only discusses parameter count and never mentions these split hazards.\n\nIf you rerun the analysis with slide-grouped splits and a fixed validation protocol (e.g., choose the backbone on a validation fold, not the test set), the relative claim that the hybrid improves over the base might survive, but the absolute numbers will come down. As is, the central empirical claim is not trustworthy. This paper is useful mainly as a cautionary example of patch-level leakage for readers working on histopathology classification.\n\nMy recommendation: don't accept this for review as submitted. If the authors redo the experiments with proper splits and a fixed model-selection protocol, the paper might be worth a fresh look. For now, the right answer is reject.","headline":"Patch-level splitting and test-set model selection undermine the headline results of this otherwise ordinary transfer-learning study; the 100% TCGA number is not evidence of slide-level generalization.","tokens_in":20485,"tokens_out":4146,"would_cite":false,"duration_ms":39951,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A hybrid convolutional network that fine-tunes the top layers of a pre-trained feature extractor and classifies through a deep fully connected stack reaches 100 percent accuracy on a public liver cancer slide dataset and 96.71 percent on…","keywords":["hepatocellular carcinoma","histopathology image classification","hybrid deep learning","transfer learning","fine-tuning","patch-based classification","H&E stained slides","convolutional neural network"],"falsifier":"Run the same hybrid training and evaluation but assign all patches from each whole-slide image (or each patient) to either the train or the test fold, never both; if development-set accuracy drops well below 100% (or hospital accuracy below 96.71%), the original random-patch split was the source of the high scores. A complementary test: take the development-trained hybrid and evaluate it on slides from a different hospital with different staining, with no fine-tuning, and compare to the reported hospital accuracy.","tokens_in":19543,"feed_emoji":"🔬","tokens_out":7012,"duration_ms":62210,"temperature":0.7,"pith_summary":"The paper claims that a simple modification to transfer learning—freezing a pre-trained CNN's lower layers, fine-tuning its upper layers, and replacing its shallow classifier with a deeper fully connected stack—consistently improves patch-level classification of hepatocellular carcinoma in H&E stained liver slides. On the public development database the best hybrid reaches 100% accuracy, sensitivity, specificity, and F1-score with AUC 1.00, while on an independent four-class hospital database the best hybrid reaches 96.71% accuracy. The same strategy also pushes three backbones to 100% on a colon histopathology dataset. If the results hold under strict data separation, the work offers a low-effort recipe for boosting cancer grading without designing new architectures.","feed_headline":"Hybrid CNN scores 100% on liver cancer grading","feed_subtitle":"Fine-tuning top layers plus a deeper classifier beats frozen baselines, reaching 96.71% on a hospital test set.","key_machinery":"The load-bearing mechanism is the hybrid model: a pre-trained convolutional feature extractor whose bottom layers stay frozen while its top layers are trainable, connected to a deep fully connected (ANN) classifier that gradually reduces feature dimensionality to the output label space. Compared with the base model—only the final layer retrained—the hybrid unfreezes a few top feature-extractor layers and adds several hidden layers. Training uses 5-fold stratified cross-validation with weighted random sampling for class balance and a cosine annealing warm-restart schedule; preprocessing includes patch extraction, a stain-vector color normalization method, and augmentation.","core_discovery":"On the paper's own terms, the central discovery is that a hybrid architecture—a pre-trained CNN feature extractor with frozen bottom layers and selectively fine-tuned top layers, attached to a custom ANN classifier with gradual dimension reduction—outperforms the corresponding base pre-trained models across every dataset tested. The gain is consistent: the minimum accuracy increase is 0.76% on the development set and 3.29% on the hospital set, with one backbone achieving perfect scores on the three-class development task and another leading on the four-class hospital task. The paper interprets this as evidence that classifier depth and selective fine-tuning matter more than which backbone is chosen, and that no single architecture is universally best across datasets.","pith_inferences":["Because the train/test split is made at the patch level rather than at the whole-slide or patient level, patches from the same slide can appear in both folds; a slide-holdout split would likely lower the reported numbers, and the 100% development-set result should be read with that caveat.","The accuracy gap between hybrid and base models may partly reflect the deeper classifier memorizing stain and tissue artifacts specific to each slide, not just tumor biology; testing on external slides from a different scanner or staining batch would separate these effects.","The stated comparison with prior work uses different patch sets and split protocols, so the headline gains over earlier methods are not a controlled comparison; a common benchmark split would be needed to verify superiority.","A practical extension would be to train the hybrid on patches but aggregate predictions at whole-slide level (e.g., majority vote), which is the clinically actionable unit and would test whether perfect patch accuracy translates to slide-level diagnosis."],"forward_implications":["The hybrid recipe can be applied to any pre-trained CNN and yields gains even when the base model already exceeds 99% accuracy, suggesting the bottleneck is classifier capacity, not feature extraction.","Near-perfect patch-level separation on the development set suggests automated triage of liver slides is feasible at patch level, though slide-level diagnosis remains untested.","The best backbone differs by dataset (ResNet50 on the development set, EfficientNetb3 on the hospital set), so practitioners should expect to select backbones per deployment site.","The same deep-classifier modification transfers to another histopathology domain (colon), indicating the strategy is not liver-specific."],"supporting_citations":[{"why":"Supplies the whole-slide images and labels used for model development.","marker":"[1]"},{"why":"Supplies the independent hospital patch dataset for validation and the baseline results the paper compares against.","marker":"[2]"},{"why":"Provides the patch-extraction and patch-level classification approach that the method builds on.","marker":"[12]"},{"why":"Defines the residual network architecture used as the best feature extractor on the development set.","marker":"[17]"},{"why":"Defines the EfficientNet scaling family from which the best hospital-set backbone is taken.","marker":"[19]"},{"why":"Motivates selective fine-tuning of top layers as a performance booster.","marker":"[25]"},{"why":"Supplies the stain-vector color normalization used in preprocessing.","marker":"[35]"}],"fun_headline_variants":["Hybrid CNN hits 100% accuracy on liver cancer grading","Hybrid deep model outperforms standard CNNs for HCC grading","Transfer learning hybrid reaches perfect score on liver cancer","Custom ANN plus CNN achieves 100% on HCC grading","Hybrid CNN boosts liver cancer grading accuracy by up to 4%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported accuracies assume that randomly dividing image patches into training and test sets keeps the sets statistically independent, meaning no patch from the same whole-slide image appears in both sets; if that leaks, the model can memorize slide-specific staining and the numbers are inflated.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid CNN hits 100% accuracy on liver cancer grading","Hybrid deep model outperforms standard CNNs for HCC grading","Transfer learning hybrid reaches perfect score on liver cancer","Custom ANN plus CNN achieves 100% on HCC grading","Hybrid CNN boosts liver cancer grading accuracy by up to 4%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001072,"raw_usage":{"total_tokens":4541,"prompt_tokens":1051,"completion_tokens":3490,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":667,"completion_tokens_details":{"reasoning_tokens":3406}},"tokens_in":667,"tokens_out":3490,"duration_ms":19852,"temperature":1.0,"reasoning_tokens":3406,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:47:20.314952+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same hybrid training and evaluation but assign all patches from each whole-slide image (or each patient) to either the train or the test fold, never both; if development-set accuracy drops well below 100% (or hospital accuracy below 96.71%), the original random-patch split was the source of the high scores. A complementary test: take the development-trained hybrid and evaluate it on slides from a different hospital with different staining, with no fine-tuning, and compare to the reported hospital accuracy.","supporting_citations":[{"cited_title":"The cancer genome atlas liver hepatocellular carcinoma collection (TCGA-LIHC),","cited_arxiv_id":null,"evidence_quote":"Supplies the whole-slide images and labels used for model development."},{"cited_title":"Livernet: efficient and robust deep learning model for automatic diagnosis of sub- types of liver hepatocellular carcinoma cancer from h&e stained liver histopathology images,","cited_arxiv_id":null,"evidence_quote":"Supplies the independent hospital patch dataset for validation and the baseline results the paper compares against."},{"cited_title":"Deep learning- based classification of liver cancer histopathology images using only global labels,","cited_arxiv_id":null,"evidence_quote":"Provides the patch-extraction and patch-level classification approach that the method builds on."},{"cited_title":"EfficientNet: Rethinking model scaling for con- volutional neural networks,","cited_arxiv_id":null,"evidence_quote":"Defines the EfficientNet scaling family from which the best hospital-set backbone is taken."},{"cited_title":"How deeply to fine-tune a convolutional neural network: A case study using a histopathology dataset,","cited_arxiv_id":null,"evidence_quote":"Motivates selective fine-tuning of top layers as a performance booster."},{"cited_title":"A method for normalizing histology slides for quantitative analysis,","cited_arxiv_id":null,"evidence_quote":"Supplies the stain-vector color normalization used in preprocessing."}],"review_version":1}