Deep learning-based compression of giga-resolution whole slide images

Erik Smistad; Etienne Gaucher; Ingerid Reinertsen; Maren H{\o}ib{\o}; Marit Valla

arxiv: 2605.17668 · v1 · pith:OQ6SZXWHnew · submitted 2026-05-17 · 💻 cs.CV

Deep learning-based compression of giga-resolution whole slide images

Maren H{\o}ib{\o} , Etienne Gaucher , Ingerid Reinertsen , Marit Valla , Erik Smistad This is my paper

Pith reviewed 2026-05-20 13:29 UTC · model grok-4.3

classification 💻 cs.CV

keywords whole slide imagesdeep learning compressiondigital pathologyimage compressionglass removalJPEG comparisonSSIM qualitystorage reduction

0 comments

The pith

Deep learning compression reduces whole slide image file sizes by 43-72 percent compared to JPEG while preserving diagnostic quality.

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

The paper examines whether deep learning tools can cut the enormous storage demands of whole slide images in digital pathology. Standard JPEG compression leaves gigabytes of data per slide, much of it empty glass background. The authors test deep learning models that first segment and remove glass regions, then compress the remaining tissue with learned codecs. When these steps are combined, total file sizes drop by 44-80 percent versus plain JPEG, with structural similarity scores staying above 0.95 on tissue patches. The work shows that learned methods outperform JPEG, JPEG-2000, and JPEG-XL on size savings while keeping visual fidelity suitable for pathology review.

Core claim

Deep learning-based tissue segmentation followed by learned compression reduces whole slide image sizes by 43-72 percent relative to JPEG alone, and by up to 80 percent when glass is also removed, while maintaining average SSIM above 0.95 on tissue patches; the same models outperform JPEG-XL and JPEG-2000 on compression ratio but require longer decompression times.

What carries the argument

Deep learning models trained for tissue segmentation to isolate glass and for end-to-end image compression on patches extracted from whole slide images.

If this is right

Deep learning compression yields 43-72 percent smaller files than JPEG on full-resolution pyramids.
Replacing glass with uniform pixels or empty tiles adds 6-62 percent extra savings depending on the codec.
On isolated tissue patches the best learned codecs save 35-40 percent over JPEG while keeping SSIM above 0.95.
JPEG-XL and JPEG-2000 achieve only 17 percent and 14 percent savings respectively on the same patches.
Decompression of the learned models takes longer than JPEG or JPEG-XL.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Pathology archives could store several times more slides in the same disk space or transmit them faster over networks.
Real-time remote consultation becomes more practical if decompression speed can be improved without losing compression gains.
The approach may extend to other high-resolution biomedical imaging modalities that contain large background regions.
A production system would still need periodic retraining or domain adaptation when scanner hardware or staining changes.

Load-bearing premise

The segmentation and compression models, trained on the authors' chosen scanners and tissue types, will continue to deliver the reported size reductions and image quality on whole slide images from different laboratories or staining protocols.

What would settle it

Measure file size and SSIM on a held-out set of whole slide images scanned with a different vendor's microscope and stained with a different protocol; if average size reduction falls below 30 percent or SSIM drops below 0.90, the claimed advantage does not generalize.

Figures

Figures reproduced from arXiv: 2605.17668 by Erik Smistad, Etienne Gaucher, Ingerid Reinertsen, Maren H{\o}ib{\o}, Marit Valla.

**Figure 1.** Figure 1: Histopathology images may include large areas of glass, here shown in gray, potentially wasting several [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Figure 2: Illustration of the proposed method. Top row: [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Creation of whole pyramids and patch pyramids. JPEG-compressed (N=63) and JPEG-XL-compressed (N=63) image pyramids were created with the original SVS structure and tile size. Three image pyramids were created with each compression format: one with glass unchanged, one with glass pixels replaced by white pixels, and one with all-glass tiles removed. The recreated JPEG and JPEG-XL pyramids were compressed wi… view at source ↗

**Figure 4.** Figure 4: Tissue segmentation with the proposed model inferred on an image of a hematoxylin and eosin [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗

**Figure 5.** Figure 5: Tissue segmentation with color thresholding and the proposed model at multiple brightness levels. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

**Figure 6.** Figure 6: Boxplot representing the size reduction (%) of new image pyramids and [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗

**Figure 7.** Figure 7: Original whole slide image (WSI) of a hematoxylin and eosin stained section from a breast tissue (top) [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗

**Figure 8.** Figure 8: Structural similarity index (SSIM), and average bits per pixel (BPP) per patch, in the balanced test set [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗

**Figure 9.** Figure 9: A patch where the trained models performed poorly. The patch is from an HE-stained slide, but is [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗

**Figure 10.** Figure 10: The JPEG-2000 and JPEG-XL patches with the lowest structural similarity index (SSIM) in the [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗

**Figure 9.** Figure 9: The [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗

**Figure 9.** Figure 9 [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗

**Figure 1-2.** Figure 1-2 [PITH_FULL_IMAGE:figures/full_fig_p025_1-2.png] view at source ↗

**Figure 3.** Figure 3: Fig.3 [PITH_FULL_IMAGE:figures/full_fig_p025_3.png] view at source ↗

**Figure 1.** Figure 1: Rate distortion curve for JPEG, JPEG-XL and the pretrained model mbt2018-mean q7 on the balanced patch dataset. SSIM against bpp. The points for JPEG and JPEG-XL are at quality factors 70, 80, and 90, whereas the deep learning-based model is at quality level 5, 6, and 7. Abbreviations: SSIM = structural similarity index, bpp = bits per pixel [PITH_FULL_IMAGE:figures/full_fig_p028_1.png] view at source ↗

**Figure 2.** Figure 2: Rate distortion curve for JPEG, JPEG-XL and the pretrained model mbt2018-mean q7 on the balanced patch dataset. PSNR against bpp. The points for JPEG and JPEG-XL are at quality factors 70, 80, and 90, whereas the deep learning-based model is at quality level 5, 6, and 7. Abbreviations: PSNR = peak signal-to-noise ratio, bpp = bits per pixel [PITH_FULL_IMAGE:figures/full_fig_p028_2.png] view at source ↗

**Figure 3.** Figure 3: Original whole slide image (WSI) of a cytokeratin 5/6 stained section from a breast cancer sample (middle) and the tissue segmentation performed with the proposed model at magnification ×2.5 with 10% overlap (right). The model segmented large parts of the tissue, but missed a considerable portion of the adipose regions. An area from the adipose region is shown in the left figure to illustrate the difficul… view at source ↗

read the original abstract

Implementation of digital pathology leads to an increased number of whole slide images (WSIs). The large size of WSIs is challenging. Today, WSIs are compressed with codecs like JPEG resulting in several gigabytes per WSI, and large amounts of space are wasted storing glass. In this study, deep learning-based tissue segmentation for glass removal, and deep learning compression methods were explored and compared with JPEG, JPEG-2000 and JPEG-XL. Image pyramids (N=21) with intact glass, glass replaced by single-colored pixels, and glass replaced by zero-byte tiles were created and compressed with JPEG, JPEG-XL and a deep learning model. Additionally, several compression models were evaluated on a tissue patch dataset and compared with JPEG, JPEG-2000 and JPEG-XL. Removing glass reduced file sizes considerably for JPEG and JPEG-XL. Deep learning-based image compression reduced the WSI size by 43-72% compared to JPEG compression, whereas deep learning-based glass removal reduced the WSI size by 0.3-33%, and 6-62% using only single-colored pixels and removing all-glass tiles, respectively. Combining the two gave a small improvement to a 44-80% total size reduction which indicates that deep learning-based image compression is able to efficiently compress glass tiles, whereas JPEG is not. On the tissue patch dataset, the best deep learning-based compression models saved on average ~35-40% per patch compared to JPEG, while keeping an average SSIM above 0.95, whereas JPEG-XL and JPEG-2000 saved 17% and 14%, respectively while keeping an SSIM of 0.96. However, the deep learning models had higher decompression times than JPEG and JPEG-XL.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This paper gets practical file size cuts on whole slide images by pairing deep learning glass removal with DL compression, but the gains rest on limited data and thin methods reporting.

read the letter

Colleague, the main thing here is that deep learning compression cut WSI file sizes by 43-72% versus JPEG on their 21 pyramids, with glass removal adding another 0.3-33% and the combination reaching 44-80% total reduction while JPEG-XL and JPEG-2000 did less on the patch set. They also kept average SSIM above 0.95 on patches. That is a usable empirical result for storage in digital pathology if the numbers travel. What the work actually does is test three versions of the pyramids (intact glass, single-color glass, zero-byte glass tiles) plus a separate tissue patch collection, then run the same DL compression models against JPEG, JPEG-2000, and JPEG-XL. The observation that DL compression handles glass tiles efficiently while JPEG does not is a clear takeaway from the numbers. They report the decompression time penalty as well, which is honest. The soft spots are straightforward. The abstract gives no architecture details, training set size, hyperparameters, or any error bars or statistical tests, so it is hard to judge how stable the 35-40% patch savings really are. Everything comes from their own scanners and protocols, so the stress-test concern about generalization to new tissue types or hardware is real and unaddressed. Decompression being slower than JPEG-XL is noted but not quantified in a way that shows whether it matters for typical workflows. This is for labs or groups already managing large WSI archives and looking for immediate storage relief without new scanners. A reader who needs concrete compression benchmarks on real pathology slides will find the direct comparisons useful. It deserves peer review because the core claims are measurable and the problem is current, though any referee will want the missing methods section and at least one external validation set before the size reductions can be treated as reliable.

Referee Report

2 major / 2 minor

Summary. The paper claims that deep learning-based tissue segmentation for glass removal and deep learning-based compression can substantially reduce WSI file sizes compared to JPEG, JPEG-2000, and JPEG-XL. On 21 image pyramids, DL compression yields 43-72% size reduction versus JPEG, glass removal yields 0.3-33% (or 6-62% with single-color/zero-byte tiles), and the combination reaches 44-80%; on a tissue patch dataset the best DL models save ~35-40% per patch at SSIM > 0.95 while JPEG-XL and JPEG-2000 save 17% and 14% at SSIM 0.96, albeit with higher decompression times.

Significance. If the reported size reductions prove robust, the work addresses a practical bottleneck in digital pathology by showing how DL methods can exploit the large glass regions and tissue statistics in WSIs to cut storage and transmission costs. The direct empirical comparisons on held-out pyramids and patches, together with concrete SSIM and file-size metrics, constitute a clear strength that could inform future codec design for giga-resolution medical images.

major comments (2)

[Methods] The manuscript provides no details on the deep learning model architectures, training procedures, hyperparameters, or the split between training and held-out test images (N=21 pyramids and patch dataset). Without this information it is impossible to assess whether the claimed 43-72% and 35-40% savings are reproducible or overfit to the authors' scanners and staining protocols.
[Results] No error bars, statistical significance tests, or cross-validation results are reported for the size-reduction figures in the abstract or results. This omission is load-bearing for the central quantitative claims, as the practical value of the 44-80% combined reduction cannot be judged without evidence of variability across images or scanners.

minor comments (2)

[Abstract] The abstract states that DL models have higher decompression times than JPEG and JPEG-XL but does not report the actual times or discuss the resulting trade-off for clinical workflows.
[Methods] Consider clarifying the exact definition of 'glass replaced by zero-byte tiles' versus 'single-colored pixels' when describing the three pyramid variants, as the 6-62% range depends on this distinction.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on reproducibility and statistical reporting. We address each major comment below and indicate the changes planned for the revised manuscript.

read point-by-point responses

Referee: [Methods] The manuscript provides no details on the deep learning model architectures, training procedures, hyperparameters, or the split between training and held-out test images (N=21 pyramids and patch dataset). Without this information it is impossible to assess whether the claimed 43-72% and 35-40% savings are reproducible or overfit to the authors' scanners and staining protocols.

Authors: We agree that the manuscript lacks sufficient methodological detail for reproducibility. In the revised version we will add a dedicated Methods section describing the deep learning architectures for compression and tissue segmentation (including network types, layer configurations, and any custom components), the training procedures and hyperparameters (optimizer, learning rate schedule, batch size, epochs, loss functions), the training datasets, and the explicit train/held-out splits for both the 21 pyramids and the tissue patch dataset. This will allow readers to evaluate potential scanner- or protocol-specific overfitting. revision: yes
Referee: [Results] No error bars, statistical significance tests, or cross-validation results are reported for the size-reduction figures in the abstract or results. This omission is load-bearing for the central quantitative claims, as the practical value of the 44-80% combined reduction cannot be judged without evidence of variability across images or scanners.

Authors: We acknowledge the absence of error bars and formal statistical tests. The reported ranges (43-72%, 44-80%) already reflect observed variation across the 21 pyramids, but we will add explicit error bars (standard deviation or interquartile range) to the size-reduction metrics in the revised results and figures. Full k-fold cross-validation on giga-resolution WSIs is computationally prohibitive; however, we will clarify the use of held-out pyramids and patches and, where feasible, report basic measures of variability or pairwise comparisons. We believe the direct empirical comparisons remain informative but will strengthen the statistical presentation as requested. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical compression results on held-out data

full rationale

The paper performs standard empirical evaluation: deep learning segmentation and compression models are trained on the authors' tissue patch and WSI datasets, then tested via direct file-size and SSIM measurements against JPEG/JPEG-XL baselines on held-out pyramids and patches. No derivation chain, equations, or predictions are presented that reduce to fitted inputs by construction. Claims of 43-72% size reduction etc. are reported outcomes of these comparisons, not self-referential definitions or self-citation load-bearing steps. The work is self-contained against external benchmarks (JPEG codecs) with no evident circularity patterns.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The paper is an empirical ML study; it relies on standard assumptions about deep learning applicability to image compression and segmentation rather than new axioms or invented entities. Free parameters consist of the many hyperparameters and architecture choices in the trained models.

free parameters (1)

Deep learning model architecture and training hyperparameters
Choices of network depth, loss functions, learning rates, and training data augmentations are tuned to produce the reported compression ratios and SSIM values.

axioms (1)

domain assumption Deep neural networks can learn domain-specific image compression that outperforms standard codecs on pathology slides while preserving diagnostic quality as measured by SSIM.
This assumption underpins the decision to explore and compare deep learning compression methods.

pith-pipeline@v0.9.0 · 5874 in / 1425 out tokens · 40772 ms · 2026-05-20T13:29:15.022061+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Deep learning-based image compression reduced the WSI size by 43-72% compared to JPEG compression, whereas deep learning-based glass removal reduced the WSI size by 0.3-33%, and 6-62% using only single-colored pixels and removing all-glass tiles, respectively.
IndisputableMonolith/Foundation/AlphaCoordinateFixation.lean J_uniquely_calibrated_via_higher_derivative unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

The models trained on pathology data were trained from scratch over 130 epochs, with the Adam optimizer and an initial learning rate of 1e-4. The loss function was a rate distortion loss

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

[1]

2022.url:https://doi.org/10.7937/DJG7-GZ87

Cancer Moonshot Biobank.Cancer Moonshot Biobank – Colorectal Cancer Collection (CMB-CRC) (Ver- sion 7) [Dataset]. 2022.url:https://doi.org/10.7937/DJG7-GZ87

work page doi:10.7937/djg7-gz87 2022
[2]

Cancer Moonshot Biobank.Cancer Moonshot Biobank – Lung Cancer Collection (CMB-LCA) (Version

work page
[3]

2022.url:https://doi.org/10.7937/3CX3-S132

[Dataset]. 2022.url:https://doi.org/10.7937/3CX3-S132

work page doi:10.7937/3cx3-s132 2022
[4]

Cancer Moonshot Biobank.Cancer Moonshot Biobank – Prostate Cancer Collection (CMB-PCA) (Version

work page
[5]

2022.url:https://doi.org/10.7937/25T7-6Y12

[Dataset]. 2022.url:https://doi.org/10.7937/25T7-6Y12

work page doi:10.7937/25t7-6y12 2022
[6]

The Cancer Imaging Archive (TCIA): maintaining and operating a public informa- tion repository

Kenneth Clark et al. “The Cancer Imaging Archive (TCIA): maintaining and operating a public informa- tion repository”. In:Journal of digital imaging26 (2013), pp. 1045–1057

work page 2013
[7]

HER2 and trastuzumab treatment response H&E slides with tumor ROI anno- tations (Version 3) [Data set]

Saman Farahmand et al. “HER2 and trastuzumab treatment response H&E slides with tumor ROI anno- tations (Version 3) [Data set]”. In: (2022).url:https://doi.org/10.7937/E65C-AM96

work page doi:10.7937/e65c-am96 2022
[8]

X. Han, J. Liu Y. abd Chen, and D Kang.Multimodal imaging of ductal carcinoma in situ with microin- vasion (HE-vs-MPM) (Version 1) [Data set]. 2023.url:https://doi.org/10.7937/3FYC-AC78

work page doi:10.7937/3fyc-ac78 2023
[9]

2020.url:https://doi.org/10.7937/TCIA.YZWQ-ZZ63

National Cancer Institute Clinical Proteomic Tumor Analysis Consortium (CPTAC).The Clinical Pro- teomic Tumor Analysis Consortium Colon Adenocarcinoma Collection (CPTAC-COAD) (Version 1) [Data set]. 2020.url:https://doi.org/10.7937/TCIA.YZWQ-ZZ63

work page doi:10.7937/tcia.yzwq-zz63 2020
[10]

2020.url:https://doi.org/10.7937/TCIA.ZS4A-JD58

National Cancer Institute Clinical Proteomic Tumor Analysis Consortium (CPTAC).The Clinical Pro- teomic Tumor Analysis Consortium Ovarian Serous Cystadenocarcinoma Collection (CPTAC-OV) (Ver- sion 3) [Data set]. 2020.url:https://doi.org/10.7937/TCIA.ZS4A-JD58

work page doi:10.7937/tcia.zs4a-jd58 2020
[11]

2018.url:https://doi.org/10.7937/K9/TCIA.2018.SC20FO18

National Cancer Institute Clinical Proteomic Tumor Analysis Consortium (CPTAC).The Clinical Pro- teomic Tumor Analysis Consortium Pancreatic Ductal Adenocarcinoma Collection (CPTAC-PDA) (Ver- sion 14) [Data set]. 2018.url:https://doi.org/10.7937/K9/TCIA.2018.SC20FO18

work page doi:10.7937/k9/tcia.2018.sc20fo18 2018
[12]

2019.url:https://doi.org/10.7937/K9/TCIA.2018.3R3JUISW

National Cancer Institute Clinical Proteomic Tumor Analysis Consortium (CPTAC).The Clinical Pro- teomic Tumor Analysis Consortium Uterine Corpus Endometrial Carcinoma Collection (CPTAC-UCEC) (Version 12) [Data set]. 2019.url:https://doi.org/10.7937/K9/TCIA.2018.3R3JUISW

work page doi:10.7937/k9/tcia.2018.3r3juisw 2019

[1] [1]

2022.url:https://doi.org/10.7937/DJG7-GZ87

Cancer Moonshot Biobank.Cancer Moonshot Biobank – Colorectal Cancer Collection (CMB-CRC) (Ver- sion 7) [Dataset]. 2022.url:https://doi.org/10.7937/DJG7-GZ87

work page doi:10.7937/djg7-gz87 2022

[2] [2]

Cancer Moonshot Biobank.Cancer Moonshot Biobank – Lung Cancer Collection (CMB-LCA) (Version

work page

[3] [3]

2022.url:https://doi.org/10.7937/3CX3-S132

[Dataset]. 2022.url:https://doi.org/10.7937/3CX3-S132

work page doi:10.7937/3cx3-s132 2022

[4] [4]

Cancer Moonshot Biobank.Cancer Moonshot Biobank – Prostate Cancer Collection (CMB-PCA) (Version

work page

[5] [5]

2022.url:https://doi.org/10.7937/25T7-6Y12

[Dataset]. 2022.url:https://doi.org/10.7937/25T7-6Y12

work page doi:10.7937/25t7-6y12 2022

[6] [6]

The Cancer Imaging Archive (TCIA): maintaining and operating a public informa- tion repository

Kenneth Clark et al. “The Cancer Imaging Archive (TCIA): maintaining and operating a public informa- tion repository”. In:Journal of digital imaging26 (2013), pp. 1045–1057

work page 2013

[7] [7]

HER2 and trastuzumab treatment response H&E slides with tumor ROI anno- tations (Version 3) [Data set]

Saman Farahmand et al. “HER2 and trastuzumab treatment response H&E slides with tumor ROI anno- tations (Version 3) [Data set]”. In: (2022).url:https://doi.org/10.7937/E65C-AM96

work page doi:10.7937/e65c-am96 2022

[8] [8]

X. Han, J. Liu Y. abd Chen, and D Kang.Multimodal imaging of ductal carcinoma in situ with microin- vasion (HE-vs-MPM) (Version 1) [Data set]. 2023.url:https://doi.org/10.7937/3FYC-AC78

work page doi:10.7937/3fyc-ac78 2023

[9] [9]

2020.url:https://doi.org/10.7937/TCIA.YZWQ-ZZ63

National Cancer Institute Clinical Proteomic Tumor Analysis Consortium (CPTAC).The Clinical Pro- teomic Tumor Analysis Consortium Colon Adenocarcinoma Collection (CPTAC-COAD) (Version 1) [Data set]. 2020.url:https://doi.org/10.7937/TCIA.YZWQ-ZZ63

work page doi:10.7937/tcia.yzwq-zz63 2020

[10] [10]

2020.url:https://doi.org/10.7937/TCIA.ZS4A-JD58

National Cancer Institute Clinical Proteomic Tumor Analysis Consortium (CPTAC).The Clinical Pro- teomic Tumor Analysis Consortium Ovarian Serous Cystadenocarcinoma Collection (CPTAC-OV) (Ver- sion 3) [Data set]. 2020.url:https://doi.org/10.7937/TCIA.ZS4A-JD58

work page doi:10.7937/tcia.zs4a-jd58 2020

[11] [11]

2018.url:https://doi.org/10.7937/K9/TCIA.2018.SC20FO18

National Cancer Institute Clinical Proteomic Tumor Analysis Consortium (CPTAC).The Clinical Pro- teomic Tumor Analysis Consortium Pancreatic Ductal Adenocarcinoma Collection (CPTAC-PDA) (Ver- sion 14) [Data set]. 2018.url:https://doi.org/10.7937/K9/TCIA.2018.SC20FO18

work page doi:10.7937/k9/tcia.2018.sc20fo18 2018

[12] [12]

2019.url:https://doi.org/10.7937/K9/TCIA.2018.3R3JUISW

National Cancer Institute Clinical Proteomic Tumor Analysis Consortium (CPTAC).The Clinical Pro- teomic Tumor Analysis Consortium Uterine Corpus Endometrial Carcinoma Collection (CPTAC-UCEC) (Version 12) [Data set]. 2019.url:https://doi.org/10.7937/K9/TCIA.2018.3R3JUISW

work page doi:10.7937/k9/tcia.2018.3r3juisw 2019