Bag-of-Visual-Words for Spatial Mapping of Lung Adenocarcinoma Growth Patterns

Darya Ardan; Henning M\"uller; Valentin Oreiller

REVIEW 4 major objections 7 minor 17 references

Bag-of-Visual-Words for Spatial Mapping of Lung Adenocarcinoma Growth Patterns

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Representing tissue regions as bags of visual words maps lung adenocarcinoma growth patterns and beats a supervised mean-pooling baseline on histologic grade classification (0.729 vs 0.678 balanced accuracy).

desk verdict A genuinely useful BoVW-at-ROI-scale mapping pipeline with honest writing, but the headline grade advantage over a supervised baseline does not survive macro-F1, and label-source leakage plus tiny N make it a hypothesis, not a result. read the letter →

arxiv 2608.05074 v1 pith:RCZUTZKN submitted 2026-08-05 cs.CV

Darya Ardan , Valentin Oreiller , Henning Müller This is my paper

classification cs.CV

keywords lungadenocarcinomagrowthpatternsbag-of-visual-wordswholeslideimagespatialmappingfoundationmodelweaklysupervisedlearninghistologicgrade

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes a weakly supervised, training-free pipeline that maps lung adenocarcinoma (LUAD) growth patterns across whole slide images by representing tissue regions as bags of visual words. The method learns a visual vocabulary by clustering tile embeddings from a frozen foundation model over 168 annotated reference regions, builds one prototype per growth pattern as the average histogram of same-label regions, and assigns every sliding-window region of a new slide to the nearest prototype under Jensen–Shannon divergence. The central claim is that this ROI-level decomposition preserves grade-relevant intratumoural heterogeneity, so for binary histologic grade classification it beats a supervised SVM trained on mean-pooled embeddings (balanced accuracy 0.729 versus 0.678 for the best encoder). If this is right, clinically useful spatial pattern maps and a competitive grade classifier can be obtained from a small reference set and no training on the evaluation cohort.

What carries the argument

The load-bearing mechanism is the ROI-level BoVW histogram. Each annotated region is encoded by assigning its tile embeddings to the nearest of $k$ K-means centroids (visual words) and normalizing the resulting one-hot counts; each growth pattern gets a prototype equal to the mean histogram of its reference regions, and unseen sliding-window regions are classified by the prototype with minimal Jensen–Shannon divergence. This converts the WSI representation from a single pooled vector into a distribution over architectural words, which is what the paper argues preserves the spatially localized high-grade signal.

What would settle it

Have a board-certified pathologist independently label the same 168 reference ROIs and recompute the prototypes and the Task 2 balanced accuracy. If the BoVW advantage over the supervised SVM disappears, or the spatial maps change materially, then the claim that ROI-level decomposition preserves grade-relevant heterogeneity would not survive corrected labels.

Watch

Extended reading notes

Core claim

The central discovery is that a bag-of-visual-words histogram computed over a region of interest, rather than a mean-pooled embedding vector, is the right unit for capturing LUAD growth patterns. With vocabulary size $k=8$ and the H-Optimus-1 encoder, the prototype classifier reaches $0.729$ balanced accuracy on binary histologic grade, above the $0.678$ of a supervised SVM on mean-pooled WSI embeddings, and the advantage holds across all three tested encoders. The paper attributes this to the fact that the ROI histogram distributes the slide's representation over pattern-relevant visual words, preserving the contribution of spatially limited high-grade regions that averaging attenuates. The same pipeline yields tumour/healthy classification at $0.974$ balanced accuracy, close to the supervised $0.987$, and produces spatially coherent pattern maps with confidence heatmaps.

Load-bearing premise

The reference ROI labels were made by a computational pathology researcher without formal pathology training, using slide metadata and published criteria, and were not verified by a pathologist; if those labels are systematically noisy, the prototypes, the grade advantage, and the spatial maps all inherit the error.

Editorial extensions

If this is right

A training-free prototype classifier built from 168 reference ROIs can match a supervised SVM for tumour/healthy detection (0.974 vs 0.987 balanced accuracy) and exceed it for grade classification (0.729 vs 0.678).
Spatial pattern maps with confidence thresholds concentrate uncertainty at pattern boundaries, so the maps can be read as a quantitative heterogeneity profile of a slide, not just a labelled image.
The task-dependent optimal vocabulary size (k=8 for grade, k=16 for tumour) implies that vocabulary granularity should be tuned to the downstream clinical task.
Because the advantage over mean-pooling holds for all three foundation-model encoders, the phenomenon is not tied to one embedding model.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

An inference the authors leave implicit is that the same ROI-level BoVW representation could serve as a feature for survival models, since the per-slide pattern histogram quantifies intratumoural heterogeneity that clinical practice currently collapses into a single dominant pattern.
A testable extension is to replace the fixed K-means vocabulary with a hierarchical vocabulary or a learned soft assignment and check whether the grade advantage grows; the paper's finding that prototype separability does not predict downstream accuracy (UNI2-h is most separable but weakest on classification) suggests vocabulary quality is not simply inter-prototype distance.
The spatial maps could be turned into a slide-level heterogeneity score, such as the Shannon entropy of pattern proportions, and that score could be tested for association with survival in the same CPTAC cohort; the paper does not run this analysis.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know two things about this paper. First, the actual contribution is real: it applies a classic BoVW pipeline to frozen foundation-model embeddings at ROI scale, giving interpretable spatial maps of LUAD growth patterns from a tiny reference set. That is a sensible and clinically motivated idea, and the tumour/healthy results (near-parity with a supervised SVM) are consistent and credible. Second, the headline claim — that BoVW beats a supervised mean-pooled SVM on histologic grade — is not established. The advantage exists for balanced accuracy only, flips to a baseline win under macro-F1 for every encoder, rests on 19 high-grade patients, has no confidence intervals or significance tests, and the best k is picked on the evaluation set.

What the paper does well: the method is described clearly, the evaluation covers three encoders and several vocabulary sizes, and the authors are upfront about the small, imbalanced reference pool and the weak grade labels. The spatial maps in Figure 2 look plausible, and the uncertainty concentrated at pattern boundaries is exactly what you would hope for. The decision to use a disjoint reference set for prototypes and evaluation is the right call.

The soft spots are proportionate. The stress-test note about label-source leakage is on target, not a manufactured worry: the reference ROIs were selected using the same CPTAC dominant-pattern metadata that defines the grade ground truth, and no pathologist checked either. If that metadata is noisy, the errors are shared between prototypes and ground truth, which can inflate the BoVW advantage. This is not direct patient leakage, but it is a real confound. The missing hyperparameters (stride, theta_min, epsilon) and the absence of code make the experiments hard to reproduce exactly; the ugly k=6 JSD=1.000 row is explained but looks like a bug or special case that should have been handled more carefully.

For whom is this paper? Someone building ROI-level pattern mapping tools in computational pathology will find the pipeline useful as a baseline or a template. The grade claim should not be cited as evidence until it is redone with a validated reference set and proper statistics.

My recommendation: send it to peer review, absolutely, but with the expectation of major revision. The idea deserves a referee, but the central claim needs to be downgraded to a hypothesis, reference labels need pathologist verification, and the evaluation needs error bars and a pre-registered k selection. If those are done, the paper could be solid.

Referee Report

4 major / 7 minor

Summary. The paper presents a weakly supervised bag-of-visual-words (BoVW) pipeline for lung adenocarcinoma (LUAD) growth-pattern mapping. A frozen foundation-model encoder embeds tiles from 168 annotated ROIs; K-means builds a visual vocabulary; per-pattern prototypes are mean BoVW histograms; sliding-window ROIs on test WSIs are assigned to the nearest prototype under Jensen-Shannon divergence; assignments are projected onto the tile grid as spatial maps. The pipeline is evaluated on 87 CPTAC-LUAD patients with three encoders and vocabulary sizes k in {6, 8, 12, 16, 18} for tumour/healthy classification and binary histologic grade classification. The best tumour/healthy balanced accuracy is 0.974 (baseline 0.987), and the best grade balanced accuracy is 0.729 (baseline 0.678), which the authors interpret as evidence that ROI-level pattern decomposition preserves grade-relevant heterogeneity that is lost by mean pooling.

Significance. If the grade advantage were statistically robust and the reference labels were independently verified, the paper would offer a useful prototype-based alternative to supervised WSI classifiers and one of the first ROI-level, pattern-specific spatial maps for LUAD. The method is simple, transparently described, and evaluated across multiple encoders and vocabulary sizes, with an honest limitations section. The main weaknesses are statistical and evidential: the central Task 2 claim rests on a small number of high-grade patients, the comparison is metric-dependent (macro-F1 favours the baseline), and the spatial maps are not validated by a pathologist.

major comments (4)

[Section 3.3, Table 4] The central claim that BoVW outperforms the supervised baseline for grade classification is not statistically supported. The largest gap (H-Optimus-1: 0.729 vs 0.678) is computed on a cohort containing only 19 high-grade patients, and the paper provides no confidence intervals, no significance test, and no correction for the screening of 15 encoder-by-k configurations. In addition, the macro-F1 column favours the baseline for all three encoders (H-Optimus-1: 0.671 vs 0.685; H0-Mini: 0.622 vs 0.650; UNI2-h: 0.641 vs 0.708), so the conclusion that BoVW preserves grade-relevant heterogeneity is metric-dependent. Please add bootstrap confidence intervals or a paired test and pre-specify the primary metric and the analysis plan.
[Section 3, reference-pool paragraph] The reference ROI labels were manually selected by a computational pathology researcher without formal pathology training using the slide metadata and published morphological criteria, and the patient-level grade labels are taken from the same CPTAC cohort. The manuscript does not establish that the prototype labels and the grade ground truth are independent; if the CPTAC dominant-pattern metadata is systematically noisy, the errors will be shared by the BoVW prototypes and the ground truth, whereas the mean-pooled SVM does not consume pattern-level labels. This is not direct patient leakage because the evaluation patients are disjoint, but it is a shared-label-source risk that can inflate the apparent BoVW advantage. Please clarify the exact provenance of the grade ground truth, have a pathologist verify at least a subset of reference ROIs, and discuss the non-independence explicitly.
[Section 3.3, Tables 3 and 4] The best vocabulary size is chosen on the evaluation set (k=16 for Task 1 and k=8 for Task 2 for H-Optimus-1), and that chosen configuration is then compared with the baseline. This is an optimistic selection procedure: the reported gap may reflect overfitting to the test set rather than a genuine property of the method. Please use nested cross-validation or a held-out validation set to select k, or report the performance of all configurations without highlighting only the best.
[Section 3.2, Figure 2] The spatial pattern maps are a stated central output, but they are evaluated only by qualitative visual inspection, and the text says a formal assessment by pathologists is left to future work. Without any pathologist review or quantitative agreement measure, the maps cannot yet be described as clinically defined pattern maps. At minimum, report a small expert review of a few WSIs or compare the BoVW maps against the ANORAK-style ROI annotations.

minor comments (7)

[Section 3.1, Table 2] The statement that minimum JSD increases monotonically with k is contradicted by Table 2: H0-Mini decreases from 0.516 at k=12 to 0.506 at k=16, and UNI2-h decreases from 0.844 at k=12 to 0.830 at k=16. Please correct the wording.
[Section 3.1, k=6] The claim that k=6 yields JSD = 1.000 by construction due to trivially disjoint prototypes is not a consequence of the label-free K-means described in Section 2.1; a label-free clustering step does not guarantee one disjoint word per pattern. Provide a derivation or revise the text.
[Section 2.4 and Section 3] The numerical values of the sliding-window stride s, the valid-tile threshold theta_min, the JSD smoothing constant epsilon, and the confidence threshold tau are not reported; please provide these hyperparameter values for reproducibility.
[Section 3, baseline description] The SVM baseline is described only as mean-pooled ROI embeddings followed by an SVM; please specify the kernel, regularization parameter, and whether embeddings were normalized, to allow exact reproduction.
[Table 4] Several rows in Table 4 are identical across different k values (e.g., H-Optimus-1 k=12, 16, 18; UNI2-h k=6 and k=12), which may indicate a rounding or reporting artifact; please clarify.
[Introduction and Related Work] The manuscript motivates against tile-level unsupervised phenotype learning methods such as HPL and PANTHER but provides no quantitative comparison; a small comparison on the same tasks would help support the claimed advantage of ROI-level BoVW.
[General] No code or data availability statement is included; please add one to support reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the BoVW prototypes are built from a disjoint reference pool and applied without fitting to evaluation labels, so the central Task 2 comparison is an external held-out evaluation, not a constructional tautology.

full rationale

The derivation chain is an evaluation, not a definitional derivation. The vocabulary is learned by k-means on tile embeddings from 168 reference ROIs (Eq. 1); prototypes are means of same-label ROI histograms (Eq. 4); WSI regions are assigned by nearest JSD (Eq. 6). None of these equations uses the evaluation ground truth. The paper explicitly states that evaluation patients are disjoint from those contributing reference ROIs and that the BoVW prototypes are applied without fitting to labels from the evaluation cohort. The central Task 2 claim is therefore an empirical comparison on held-out patients, not a quantity equal by construction to its input. The acknowledged limitations—small imbalanced ROI pool, weak patient-level grade labels, no pathologist verification of reference ROI labels, and use of CPTAC dominant-pattern metadata to select reference ROIs—are validity and label-noise concerns, not circularity: they do not make the reported balanced accuracy equivalent to an input by equation. There are no self-citations, imported uniqueness theorems, or ansatz smuggling via citation. The k=6 JSD=1.000 degeneracy is explicitly excluded and not load-bearing. Consequently, no circular step meeting the evidence standard is present.

Assumptions & free parameters 5 free parameters · 6 assumptions · 2 invented entities

The evaluation rests on assumptions the paper does not independently verify: frozen-embedding suitability for LUAD patterns, the ROI field of view as the correct unit (carried by a single arXiv citation), non-pathologist-curated reference labels, and CPTAC metadata as ground truth for both reference patterns and grade labels. The method contributes hand-chosen thresholds (delta, tau) and a scanned vocabulary size k; no constants are fitted to the evaluation cohort, which keeps the circularity burden low.

free parameters (5)

vocabulary size k = best per task on evaluation cohort: 16 (Task 1), 8 (Task 2)
Swept over {6,8,12,16,18}; headline numbers use the evaluation-cohort-best k with no validation split, and Task 1 and Task 2 prefer different k.
prototype separability threshold delta = 0.05
Hand-set quality criterion in Section 2.3; inert because no configuration was excluded by it.
confidence threshold tau = 0.5
Hand-set in Section 2.4; controls Unknown regions and neighbour relabelling in the spatial maps.
JSD smoothing constant epsilon = unspecified
Applied in Section 2.3 so JSD is defined for zero-count bins; the value is not reported and affects prototype distances and nearest-prototype retrieval.
sliding-window stride s and valid-tile threshold theta_min = unspecified
Both defined in Section 2.4 but never given numeric values; they determine which windows are encoded and their overlap, so exact reproduction is impossible without them.

assumptions (6)

domain assumption Frozen foundation-model embeddings of 224x224 tiles encode signal sufficient to separate the six LUAD pattern classes at ROI level.
Invoked in Section 2.1 for vocabulary learning; validated only indirectly through downstream accuracy and inter-prototype JSD.
domain assumption A 1344x1344 px ROI (36 tiles) is the right unit because growth patterns are architectural phenomena spanning multiple tile widths.
Stated in the Introduction and Section 2, supported by citation [7] and annotation protocols; if false, the whole sliding-window design is misframed.
domain assumption The 168 reference ROI labels, manually selected by a researcher without formal pathology training from CPTAC dominant-pattern metadata, are accurate enough to build prototypes.
Section 3 describes the curation; no pathologist verification is reported, and the paper discloses this in Limitations.
domain assumption CPTAC patient-level binary grade labels are a valid ground truth for evaluating highest-grade-wins prediction.
Section 3.3 and Limitations state labels are weak and no WSI-level ground truth exists; the grade labels and the reference pattern labels both derive from the same pattern-metadata source.
ad hoc to paper Nearest-prototype retrieval under JSD over a K-means codebook of 6 to 18 words is a sufficient representation for pattern mapping.
Equations (1)-(6) define the scheme; no theoretical guarantee is offered, and the k=6 degeneracy in Table 2 shows the representation is fragile at small k.
domain assumption A WSI is tumour if any ROI is assigned a cancer pattern; single false positives are tolerated by the reported precision.
Task 1 decision rule in Section 2.4; the rule is crude (one false-positive ROI flips a slide) and its validity is shown only empirically by the high precision values in Table 3.

invented entities (2)

Visual vocabulary of k centroids (visual words)
purpose: Codebook for BoVW histogram encoding of tiles (Eq. 1)
Algorithmic construct from K-means on embeddings; the paper makes no claim that the words correspond to biological entities.
Per-pattern prototypes mu_g as mean BoVW histograms
purpose: Nearest-prototype assignment under JSD (Eqs. 4-6)
Statistical summaries of labelled ROIs; no independent falsifiable handle outside the pipeline.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Bag-of-Visual-Words for Spatial Mapping of Lung Adenocarcinoma Growth Patterns." pith.science (2026). https://pith.science/paper/RCZUTZKN

@misc{pith2026260805074,
  author       = {Pith},
  title        = {Pith review of: Bag-of-Visual-Words for Spatial Mapping of Lung Adenocarcinoma Growth Patterns},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RCZUTZKN}},
  note         = {Machine review of arXiv:2608.05074}
}

abstract

Spatial mapping of lung adenocarcinoma (LUAD) growth patterns across whole slide images (WSIs) requires resolving architectural context at the region level, yet existing methods operate at the individual tile level and produce generic morphological clusters rather than clinically defined pattern maps. We propose a weakly supervised Bag-of-Visual-Words (BoVW) pipeline that learns a visual vocabulary from frozen foundation model embeddings extracted from a small set of annotated regions of interest (ROIs). Pattern prototypes are constructed as mean BoVW histograms of same-label ROIs and used for nearest-prototype classification of sliding-window regions under Jensen--Shannon divergence. The resulting predictions are projected onto the WSI tile grid to produce interpretable spatial pattern maps. We evaluate the method on 87 CPTAC-LUAD patients using three foundation model encoders and multiple vocabulary sizes on two clinically motivated tasks. For tumour/healthy classification, the best configuration achieves a balanced accuracy of $0.974$ with H-Optimus-1, approaching the $0.987$ obtained by a supervised SVM trained on mean-pooled WSI embeddings. For binary histologic grade classification, the BoVW pipeline achieves higher balanced accuracy than the supervised baseline for all encoders, suggesting that ROI-level pattern decomposition preserves grade-relevant heterogeneity that is attenuated by global mean pooling.

Figures

Figures reproduced from arXiv: 2608.05074 by the authors.

**Figure 1.** Overview of the proposed pipeline. (I, II): Visual vocabulary learning from annotated ROIs via a frozen foundation model encoder, and BoVW histogram encoding. (III): Sliding-window BoVW encoding of WSIs and nearest-prototype classification under JSD. (IV): Downstream outputs: tumour/healthy classification, histologic grade classification, and spatial LUAD pattern maps. predictions are assembled into spatial patter… view at source ↗

**Figure 2.** Spatial pattern maps for two CPTAC-LUAD WSIs (H-Optimus-1, k=8). Pattern overlay (left) shows per-ROI assignments superimposed on the tissue thumbnail. Confidence heatmap (right) encodes classification confidence, with yellow indicating high confidence and dark purple indicating low confidence or Unknown regions. consistent advantage across encoders demonstrates that distributing WSI representation over pattern-re… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 15 canonical work pages

[1]

Moreira, A.L., et al.: A grading system for invasive pulmonary adenocarcinoma: a proposal from the International Association for the Study of Lung Cancer Pathol- ogy Committee.J. Thorac. Oncol.15(10), 1599–1610 (2020)

work page 2020
[2]

Pan, X., AbdulJabbar, K., Coelho-Lima, J., et al.: The artificial intelligence-based model ANORAK improves histopathological grading of lung adenocarcinoma.Na- ture Cancer5(2), 347–363 (2024)

work page 2024
[3]

Rep.9, 3358 (2019)

Wei, J.W., Tafe, L.J., Linnik, Y.A., Vaickus, L.J., Tomita, N., Hassanpour, S.: Pathologist-level classification of histologic patterns on resected lung adenocarci- noma slides with deep neural networks.Sci. Rep.9, 3358 (2019)

work page 2019
[4]

Lami, K., et al.: Standardized classification of lung adenocarcinoma subtypes and improvement of grading assessment through deep learning.Am. J. Pathol.193(12), 2066–2079 (2023)

work page 2023
[5]

Quiros, A.C., Coudray, N., Yeaton, A., et al.: Mapping the landscape of histo- morphological cancer phenotypes using self-supervised learning on unannotated pathology slides.Nature Commun.15, 4596 (2024)

work page 2024
[6]

arXiv:2405.11643 (2024)

Song, A.H., Chen, R.J., Ding, T., et al.: Morphological Prototyping for Unsupervised Slide Representation Learning in Computational Pathology. arXiv:2405.11643 (2024)

arXiv 2024
[7]

CellOMaps: A Compact Representation for Robust Classification of Lung Adenocarcinoma Growth Patterns

Reisenbüchler, D., et al.: CellOMaps: A Compact Representation for Robust Clas- sification of Lung Adenocarcinoma Growth Patterns. arXiv:2501.08094 (2025)

work page Pith review arXiv 2025
[8]

In: ICCV, pp

Sivic,J.,Zisserman,A.:VideoGoogle:Atextretrievalapproachtoobjectmatching in videos. In: ICCV, pp. 1470–1477 (2003)

work page 2003

Show all 17 references

[9]

In: AIME, pp

Caicedo, J.C., Cruz, A., Gonzalez, F.A.: Histopathology image classification using bag of features and kernel functions. In: AIME, pp. 126–135 (2009)

2009
[10]

Travis, W.D., et al.: The 2015 World Health Organization Classification of Lung Tumors.J. Thorac. Oncol.10(9), 1243–1260 (2015)

2015
[11]

Sica, G., et al.: A grading system of lung adenocarcinomas based on histologic pattern is predictive of disease recurrence in stage I tumors.Am. J. Surg. Pathol. 34(8), 1155–1162 (2010)

2010
[12]

Proteome Res.14(6), 2707–2713 (2015)

Edwards, N.J., et al.: The CPTAC Data Portal: A Resource for Cancer Proteomics Research.J. Proteome Res.14(6), 2707–2713 (2015)

2015
[13]

In: Proc

Scalbert, M., Saillard, C., Peeters, T., et al.: H-optimus-1: A foundation model for computational histopathology. In: Proc. AACR Annual Meeting, Cancer Research 86(8_Suppl), LB174 (2026)

2026
[14]

arXiv:2501.16239 (2025)

Filiot, A., Dop, N., Tchita, O., et al.: Distilling foundation models for robust and efficient models in digital pathology. arXiv:2501.16239 (2025)

2025 arXiv
[15]

Chen, R.J., Ding, T., Lu, M.Y., et al.: Towards a general-purpose foundation model for computational pathology.Nature Medicine(2024)

2024
[16]

In: ICML, pp

Ilse,M.,Tomczak,J.,Welling,M.:Attention-baseddeepmultipleinstancelearning. In: ICML, pp. 2127–2136 (2018)

2018
[17]

Pathol.25(12), 1574–1583 (2012)

Thunnissen, E., et al.: Reproducibility of histopathological subtypes and invasion in pulmonary adenocarcinoma.Mod. Pathol.25(12), 1574–1583 (2012)

2012

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Moreira, A.L., et al.: A grading system for invasive pulmonary adenocarcinoma: a proposal from the International Association for the Study of Lung Cancer Pathol- ogy Committee.J. Thorac. Oncol.15(10), 1599–1610 (2020)

work page 2020

[2] [2]

Pan, X., AbdulJabbar, K., Coelho-Lima, J., et al.: The artificial intelligence-based model ANORAK improves histopathological grading of lung adenocarcinoma.Na- ture Cancer5(2), 347–363 (2024)

work page 2024

[3] [3]

Rep.9, 3358 (2019)

Wei, J.W., Tafe, L.J., Linnik, Y.A., Vaickus, L.J., Tomita, N., Hassanpour, S.: Pathologist-level classification of histologic patterns on resected lung adenocarci- noma slides with deep neural networks.Sci. Rep.9, 3358 (2019)

work page 2019

[4] [4]

Lami, K., et al.: Standardized classification of lung adenocarcinoma subtypes and improvement of grading assessment through deep learning.Am. J. Pathol.193(12), 2066–2079 (2023)

work page 2023

[5] [5]

Quiros, A.C., Coudray, N., Yeaton, A., et al.: Mapping the landscape of histo- morphological cancer phenotypes using self-supervised learning on unannotated pathology slides.Nature Commun.15, 4596 (2024)

work page 2024

[6] [6]

arXiv:2405.11643 (2024)

Song, A.H., Chen, R.J., Ding, T., et al.: Morphological Prototyping for Unsupervised Slide Representation Learning in Computational Pathology. arXiv:2405.11643 (2024)

arXiv 2024

[7] [7]

CellOMaps: A Compact Representation for Robust Classification of Lung Adenocarcinoma Growth Patterns

Reisenbüchler, D., et al.: CellOMaps: A Compact Representation for Robust Clas- sification of Lung Adenocarcinoma Growth Patterns. arXiv:2501.08094 (2025)

work page Pith review arXiv 2025

[8] [8]

In: ICCV, pp

Sivic,J.,Zisserman,A.:VideoGoogle:Atextretrievalapproachtoobjectmatching in videos. In: ICCV, pp. 1470–1477 (2003)

work page 2003