{"id":"04b32dfb-b806-433e-b50c-521b4f79db77","arxiv_id":"2411.08530","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Selecting a small set of high-cellularity patches and encoding them with Fisher vectors gives whole-slide classification accuracy comparable to processing all patches, with much lower compute.","lead":"This paper proposes a cheaper way to classify whole-slide pathology images: keep only the most cell-dense patches and represent them with a compact statistical encoding instead of processing every region. If it works, automated cancer diagnosis on large tissue scans could run on far less computing power while staying roughly as accurate.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The nucleus-count patch selection is the load-bearing premise, yet the paper's own Camelyon17 ablation shows random patches (0.81) beat high-cellularity patches (0.80), directly contradicting the 'focused analysis' claim.","rationale":"The reader's verdict is REJECT, based largely on the two weakest assumptions listed: the nucleus-count selection premise and the un-fit GMM. I identify the nucleus-count selection as the single most load-bearing concern because it directly supports the 'focused analysis of select patches' portion of the central claim, and the paper's own data contradict it. The Fisher vector parameter issue (hand-set π=0.2, σ=0.1 without fitting) is also serious, but even if that were fixed, the selection heuristic still lacks evidence; indeed, random patches performing comparably or better would render the method's main novelty moot. The paper's internal ablation on Camelyon17 is a controlled comparison (same fixed number of patches, same training protocol) and shows random slightly beating selective patches. This is not a matter of external consensus; it is an internal inconsistency between the claimed benefit of nucleus-based selection and the reported results. A rigorous test would isolate the selection variable and determine whether the heuristic contributes anything. Since this concern reinforces the reader's REJECT rather than changing it, I set verdict_should_be to UNCHANGED and agreement to 'agree' because the reader identified the same load-bearing assumption as a key weakness.","tokens_in":8640,"tokens_out":3099,"duration_ms":29116,"concrete_test":"Run the identical classification pipeline (same ResNet-50 backbone, same fixed patch count, same Fisher vector encoding with the paper's hand-set parameters, same training scheme) on Camelyon17 and TCGA Lung, comparing HistomicsTK high-nucleus-count patch selection against uniform random patch selection, repeated over at least 5 random seeds and reporting mean and variance of accuracy/AUC. If random selection matches or exceeds nucleus-count selection on either dataset, the nucleus-count heuristic is not load-bearing and the central claim of 'focused analysis of select patches' fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that analyzing 'select patches' based on pathological significance, encoded via Fisher vectors, matches or surpasses standard practices while reducing compute. The selection mechanism is the nucleus-count heuristic from HistomicsTK (Section 3: 'get the patches with high nuclei count'). This premise is what distinguishes the method from random-sampling baselines. However, the Camelyon17 ablation in Section 5 reports that a randomly chosen fixed set of patches yields 0.81 accuracy, while high-cellularity patches yield only 0.80. The paper itself concedes: 'regions of high cellularity may not serve as reliable morphological biomarkers for metastasis detection.' Thus, on the primary dataset where selection is tested cleanly, the proposed selective mechanism does not improve accuracy over random selection. The only evidence favoring selection is the TCGA Lung ablation (random 0.67 vs. high-cellularity 0.85), but that comparison uses ResNet-50 while the high-accuracy MobileNetV3small result (0.80/0.80) uses 'a larger number of fixed patches,' confounding the selection criterion with patch count and backbone. Without a controlled comparison that fixes patch count, backbone, and all training details, the TCGA result could be attributable to those other factors. The load-bearing assumption—that nucleus count identifies diagnostically informative patches—is therefore not established and is actively contradicted by the Camelyon17 ablation. This undermines the qualitative claim that 'focused analysis of select patches' is responsible for the reported performance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage whole slide image (WSI) classification method: (i) preprocess each slide to select a fixed number of high-nucleus-count patches using HistomicsTK, and (ii) encode deep features from these patches with a Fisher vector (FV) representation built from a five-component Gaussian mixture model, followed by average pooling and a small fully connected classifier. The authors evaluate on TCGA Lung EGFR mutation classification and Camelyon17 metastasis detection, comparing against the DFVC baseline [1] and a multiple-instance-learning approach [3]. They claim that the focused selection of informative patches combined with FV encoding matches or exceeds standard practices while reducing computational load. The paper includes ablations comparing high-cellularity patch selection with random patch selection, and experiments with ResNet-50 and MobileNetV3small backbones.","tokens_in":8952,"tokens_out":4633,"duration_ms":31908,"significance":"If the claims were established, the method would provide a computationally efficient and clinically relevant alternative to exhaustive WSI processing, and the preprocessing pipeline (artifact removal, tissue detection, patching, nucleus counting) is a useful practical contribution. The paper also attempts ablations and multiple backbones, which is commendable. However, the central claims are not supported by the experiments as reported: the Fisher vector is not constructed according to the theory (GMM parameters are not fitted), the nucleus-count selection is contradicted by the Camelyon17 ablation, the TCGA ablation is confounded, and no uncertainty estimates are provided. The manuscript does not provide code or detailed experimental protocols, further limiting reproducibility. These issues are load-bearing rather than cosmetic, so the current results cannot be taken as evidence for the method's effectiveness.","major_comments":[{"comment":"The Fisher vector is not properly defined and is not fitted to the data. The text states that for a GMM with N Gaussians, the FV dimensionality is 2ND, but Eq. (1) claims FV: R(D+m) → R2m and displays a vector with 2M blocks. More importantly, the gradient formulas (2)-(3) require means and variances of a GMM estimated from the descriptor distribution, yet Section 4 states that the number of centers is set to five and that π_m=0.2 and σ_m=0.1 are configured by hand. No GMM fitting step is described anywhere. With arbitrary mixture parameters, the encoding is not a Fisher vector in the standard sense and does not capture the distribution of descriptors; the paper's claim that FV 'captures fine-grained details' is therefore unsupported. This is central to the method's novelty.","section":"Section 3, Eqs. (1)-(4), and Section 4"},{"comment":"The Camelyon17 ablation contradicts the central selection premise. The text reports that high-cellularity fixed patches yield 0.80 accuracy, while randomly chosen fixed patches yield 0.81 accuracy, and the authors themselves state that 'regions of high cellularity may not serve as reliable morphological biomarkers for metastasis detection.' Since nucleus-count selection is the core of the proposed focused-analysis approach, this result directly undermines the Abstract's claim that focused analysis of select patches aligns with or surpasses standard practices. No statistical significance test or repeated-run analysis is provided, so the 0.01 difference may be noise, but regardless, the proposed selection mechanism shows no benefit on this dataset.","section":"Section 5, Table 2 and accompanying text"},{"comment":"The TCGA Lung ablation that favors high-cellularity selection is confounded. The random-patch ablation uses ResNet-50 and reports 0.67 accuracy, while the MobileNetV3small result of 0.80 accuracy is described as using 'a larger number of fixed patches.' The comparison does not hold fixed the backbone, the number of patches, or other training details, so the difference between random and high-cellularity selection cannot be attributed to the selection criterion itself. A controlled experiment varying only the selection method, with identical patch count, backbone, and training configuration, is required to support the claim.","section":"Section 5, Table 1 and ablation paragraph"},{"comment":"No error bars, standard deviations, confidence intervals, or number of random seeds are reported for any experimental result. Each number appears to be from a single run. Given the small dataset sizes (159 TCGA slides; 500 Camelyon17 slides) and the known variability of deep-learning-based WSI classification, the reported differences (e.g., 0.80 vs. 0.81 on Camelyon17) are well within plausible run-to-run variation. Without repeated runs or uncertainty quantification, the comparative claims are not reliable.","section":"Section 5, Tables 1 and 2"}],"minor_comments":[{"comment":"The notation is ambiguous: the text reports 'accuracy of 0.85%' and 'AUC of 0.73%' while the tables are in units of percentage points (e.g., 0.85 accuracy would be 85%). This should be clarified to avoid confusion.","section":"Abstract and Section 5"},{"comment":"The introduction contains an unresolved reference 'Fig. ??' for the sample images of the datasets; this should be fixed.","section":"Section 1"},{"comment":"There are typos such as 'dimenssion' and 'prepossessing' that should be corrected.","section":"Section 3"},{"comment":"The text reports 'experiments utilizing the more compact MobileNetV3small model resulted in an accuracy of 0.74% and an AUC of 0.75%' while Table 2 lists 0.72 accuracy and 0.75 AUC for MobileNetV3small on Camelyon17; these numbers are inconsistent.","section":"Section 5, Table 2 versus text"},{"comment":"Reference [25] is incorrectly formatted ('S Jian, H Kaiming, R Shaoqing, and Z Xiangyu') and should be He et al., Deep Residual Learning for Image Recognition; please verify all reference entries.","section":"References"}],"recommendation":"reject","confidential_remarks":"The paper's central contribution, nucleus-count patch selection, is contradicted by its own clean ablation on Camelyon17, and the Fisher vector encoding is not implemented as a genuine Fisher vector because the GMM is never fitted. These are load-bearing issues that cannot be fixed by minor edits; a substantially different experimental design and likely a different selection mechanism would be needed. The use of self-citations [3] and [21] is acceptable as they provide baselines and preprocessing tools, but the [3] comparison lacks experimental detail. I would not encourage resubmission in the current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is DFVC plus a nucleus-count patch preselector, benchmarked on TCGA Lung and Camelyon17. The Fisher vector machinery is standard; the only genuinely new ingredient is selecting high-nucleus-count patches before encoding, and the paper's own Camelyon17 ablation says that selection does not beat random patches (0.80 vs 0.81). That undercuts the abstract's central claim.\n\nWhat the paper does well: it is an honest, readable engineering report. The preprocessing pipeline (artifact removal, tissue detection, filtering, nucleus counting) is sensible and reusable. The authors include ablations and openly report the Camelyon17 result against selection, conceding that high cellularity may not be a reliable biomarker for metastasis. The efficiency direction—small fixed patch set plus compact Fisher vector—is plausible and worth exploring. The citation pattern is fine: self-citations supply a preprocessing tool and a comparison number, not a circular definition of the central claim.\n\nThe soft spots are real and load-bearing. Eq. (1) declares the FV dimension as 2m while the text says 2ND; either way the standard FV has 2D dimensions per Gaussian, so this is at least a typo, but combined with the hand-set π_m=0.2 and σ_m=0.1 it tells you the GMM is never fit to the feature distribution. That makes the encoding a fixed nonlinear projection rather than a proper Fisher vector. No seeds, folds, or error bars appear in Tables 1–2, so the reported differences (e.g., 0.85 vs 0.80 accuracy on TCGA) are point estimates without uncertainty. The TCGA ablation that favors selection is confounded: it compares ResNet-50 with a fixed patch set against MobileNetV3small with a larger patch set, so you cannot attribute the gain to selection. And versus the DFVC baseline, the improvements are inconsistent: on TCGA ResNet-50 the paper gets higher accuracy but lower AUC (0.725 vs 0.77). \"Surpasses standard practices\" overstates a single mixed result.\n\nOverall: the paper is a legitimate incremental extension with some useful engineering, but the central premise is not established and the evaluation is too thin for acceptance. It is not incoherent or dishonest; a serious referee could usefully push the authors to fix the dimension error, fit the GMM, report variance, and run a controlled patch-count/backbone-matched ablation. I would send it out rather than desk-reject, but I would expect heavy revision.","headline":"DFVC plus nucleus-count patch selection, honestly reported but undercut by its own Camelyon ablation; worth a referee round with expectation of heavy revision.","tokens_in":9518,"tokens_out":2726,"would_cite":false,"duration_ms":23381,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T45","68T07","92C50"],"pacs":[],"model":"deepseek-v4-flash","headline":"A small set of nucleus-rich patches plus a Fisher vector can classify whole slide images as accurately as exhaustive analysis.","keywords":["whole slide image classification","Fisher vector","patch selection","digital pathology","weakly supervised learning","nucleus counting","computational pathology","EGFR mutation prediction"],"falsifier":"Compare the proposed pipeline against an identical pipeline in which the fixed patch set is chosen at random instead of by nucleus count across a broad range of WSI tasks; a statistically powered replication where random selection consistently matches or beats nucleus-count selection would falsify the claim that nucleus-rich patches are the informative ones. Additionally, fit the GMM parameters by EM on the patch features instead of fixing $\\pi_m=0.2$ and $\\sigma_m=0.1$; if fitted parameters never improve over fixed ones, the FV codebook's learned statistics are not doing the work claimed.","tokens_in":8384,"feed_emoji":"🔬","tokens_out":6424,"duration_ms":56613,"temperature":0.7,"pith_summary":"This paper tries to establish that a whole slide image can be classified from a small, fixed set of diagnostically selected patches encoded as a Fisher vector, rather than by exhaustively processing every patch. The authors argue that this focused patch selection plus Fisher vector representation matches or exceeds the accuracy of established full-slide and multiple-instance-learning methods while substantially reducing computation and memory. The claim is tested on EGFR mutation prediction in the TCGA Lung dataset and breast cancer metastasis detection in the Camelyon17 dataset. If correct, the result would make WSI classification more scalable for routine digital pathology workloads.","feed_headline":"A few nucleus-rich patches match whole-slide cancer classification","feed_subtitle":"Fisher-vector coding of selected patches cuts compute while keeping accuracy on lung and lymph-node slide datasets.","key_machinery":"The central object is the Fisher vector representation of patch features, together with the nucleus-count patch selector. A Fisher vector encodes the distribution of a set of $D$-dimensional patch features by computing gradients of the log-likelihood with respect to the parameters of a Gaussian mixture model with $N$ Gaussians; concatenating the mean and covariance gradient blocks gives a vector of dimension $2ND$. The paper uses $N=5$, hand-set $\\pi_m=0.2$ and $\\sigma_m=0.1$, and compresses backbone features to $D=10$ dimensions, so the final per-slide descriptor is compact. The patch selector uses HistomicsTK tissue segmentation and nucleus counting to pick a fixed number of $512\\times512$ patches at 40x magnification from the tissue region. This mechanism carries the argument by converting a WSI with hundreds of thousands of patches into a single small vector that a small classifier can handle.","core_discovery":"The paper's central claim is that a whole slide image can be summarized by a compact Fisher vector computed from a small, fixed set of diagnostically selected patches, and that this summary supports classification that matches or exceeds exhaustive patch processing. The pipeline removes artifacts, detects tissue with HistomicsTK, extracts $512\\times512$ patches at 40x zoom, filters noisy patches, and selects a fixed number of patches with the highest nucleus count. Features from a ResNet-50 or MobileNetV3small backbone are compressed to 10 dimensions, encoded through a Fisher vector with five Gaussians using fixed mixture weights $\\pi_m=0.2$ and variances $\\sigma_m=0.1$, and then average-pooled and passed to linear classifier layers. On the TCGA Lung EGFR binary task the ResNet-50 version reaches 0.85 accuracy (0.725 AUC), above the 0.80 accuracy of the deep Fisher vector coding baseline; on Camelyon17 it reaches 0.80 accuracy (0.72 AUC), above the baseline's 0.75. The paper also reports that its Camelyon17 ablation with randomly chosen fixed patches reached 0.81 accuracy, slightly above the 0.80 from high-cellularity patches, while the TCGA ablation favored high-cellularity patches (0.85 versus 0.67 random).","pith_inferences":["A testable extension: if the fixed-parameter Fisher vector performs as reported, a fitted GMM codebook may be unnecessary for small patch sets; comparing EM-fitted parameters against the fixed $\\pi_m=0.2$, $\\sigma_m=0.1$ settings would settle whether the codebook statistics actually matter.","Because random patches matched nucleus-count patches on Camelyon17, the informative-patch selection step may need to be task-specific or learned rather than based solely on cellularity.","Fisher-vector slide encoding with only five Gaussians produces a very compact descriptor, which suggests the same representation could serve fast slide-level retrieval or cross-center domain adaptation where per-slide storage and transfer cost matter."],"forward_implications":["A whole slide image can be classified from a small, fixed number of patches rather than every patch in the tissue region.","Fisher vector encoding of the selected patches yields accuracy on the TCGA Lung EGFR task and the Camelyon17 metastasis task that is comparable to or better than the deep Fisher vector coding baseline.","Because only selected patches are processed, memory and compute per slide drop, making the pipeline more feasible for routine clinical workloads.","The method is end-to-end trainable from patch selection through classifier, unlike MIL pipelines that keep feature extraction separate."],"supporting_citations":[{"why":"Supplies the deep Fisher vector coding baseline and the patch-feature descriptor convention the method builds on.","marker":"[1]"},{"why":"Provides the HistomicsTK tissue detection and nucleus counting used to select informative patches.","marker":"[22]"},{"why":"Provides the TCGA lung slides and EGFR mutation labels used for the binary classification benchmark.","marker":"[2]"},{"why":"Provides the Camelyon17 lymph node slides and metastasis labels used for the second benchmark.","marker":"[4]"},{"why":"Supplies the Fisher vector framework that motivates encoding patch feature distributions.","marker":"[20]"},{"why":"Provides the MIL patch-classification baseline whose poor result at fixed patch counts motivates the FV pipeline.","marker":"[3]"},{"why":"Supplies the GMM gradient formulation used to define the Fisher vector encoding equations.","marker":"[23]"},{"why":"Justifies the $2ND$ dimensionality statement for Fisher vectors with $N$ Gaussians and $D$-dimensional features.","marker":"[24]"}],"fun_headline_variants":["Fisher vectors from few patches match whole-slide accuracy","Sparse patch selection with Fisher vectors cuts compute","Whole-slide classification from just a few patches","Patch-sparse Fisher coding rivals exhaustive slides","A handful of patches benchmark full-slide diagnosis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a patch's nucleus count marks its diagnostic value for every target task; the paper's own Camelyon17 ablation, where random fixed patches scored 0.81 versus 0.80 for high-cellularity patches, suggests this premise may not hold there.","fun_headline_variants_meta":{"raw":{"variants":["Fisher vectors from few patches match whole-slide accuracy","Sparse patch selection with Fisher vectors cuts compute","Whole-slide classification from just a few patches","Patch-sparse Fisher coding rivals exhaustive slides","A handful of patches benchmark full-slide diagnosis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0003,"raw_usage":{"total_tokens":1781,"prompt_tokens":1045,"completion_tokens":736,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":661,"completion_tokens_details":{"reasoning_tokens":665}},"tokens_in":661,"tokens_out":736,"duration_ms":7320,"temperature":1.0,"reasoning_tokens":665,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T21:31:50.689502+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the proposed pipeline against an identical pipeline in which the fixed patch set is chosen at random instead of by nucleus count across a broad range of WSI tasks; a statistically powered replication where random selection consistently matches or beats nucleus-count selection would falsify the claim that nucleus-rich patches are the informative ones. Additionally, fit the GMM parameters by EM on the patch features instead of fixing $\\pi_m=0.2$ and $\\sigma_m=0.1$; if fitted parameters never improve over fixed ones, the FV codebook's learned statistics are not doing the work claimed.","supporting_citations":[{"cited_title":"Deep fisher vector coding for whole slide image classification","cited_arxiv_id":null,"evidence_quote":"Supplies the deep Fisher vector coding baseline and the patch-feature descriptor convention the method builds on."},{"cited_title":"Improving the fisher kernel for large-scale image classification","cited_arxiv_id":null,"evidence_quote":"Supplies the Fisher vector framework that motivates encoding patch feature distributions."},{"cited_title":"Enhanced bag of visual words representations for content based image retrieval: a comparative study","cited_arxiv_id":null,"evidence_quote":"Supplies the GMM gradient formulation used to define the Fisher vector encoding equations."},{"cited_title":"Understanding the Fisher Vector: a multimodal part model","cited_arxiv_id":"1504.04763","evidence_quote":"Justifies the $2ND$ dimensionality statement for Fisher vectors with $N$ Gaussians and $D$-dimensional features."}],"review_version":1}