资源报告

CRC-MSI 资源报告

官方名称 TCGA-CRC-DX · 38 个字段,每个字段附来源原文 · 来源核验于 2026 年 5–6 月

CRC-MSI 在 Zenodo 上的正式名称为 TCGA-CRC-DX(记录 3832231,题名 “Histological image tiles for TCGA-CRC-DX, color-normalized, sorted by MSI status, train/test split”)。它从 TCGA 结直肠癌 H&E 全切片的肿瘤区域裁出图像块,用于微卫星不稳定性(MSI)状态分类,公开为 TRAIN.zip 与 TEST.zip 两个压缩包。图像块边长 256 μm,保存为 512 × 512 像素(0.5 μm/像素),经颜色归一化;每个图像块继承所属患者的 MSI 状态。按压缩包文件清单统计,共 423 名患者、428 张切片、51,918 个图像块。

基础档案

数据集名称

TCGA-CRC-DX

Zenodo 题名与官方代码仓库 DeepHistology 的说明文件都使用 TCGA-CRC-DX。CRC-MSI 是按 MSI 任务使用的别名,不是官方名称。

来源Zenodo 记录 3832231 · 标题/页面标题

"Histological image tiles for TCGA-CRC-DX, color-normalized, sorted by MSI status, train/test split"

资源类型

数据集

Zenodo 将其登记为数据集;没有挑战赛或基准所具有的提交入口、排行榜、隐藏测试标签或评测协议。

来源Zenodo API record 3832231 · metadata.resource_type / status

"resource_type": {"title": "Dataset", "type": "dataset"}"; ""status": "published""

发布日期

2020-05

取 Zenodo 数据发布日期 2020-05-18。相关论文于 2020 年 10 月刊出,不作为数据发布时间。

来源Zenodo API record 3832231 · metadata.publication_date

"publication_date": "2020-05-18""

数据集主页

DOI 10.5281/zenodo.3832231 指向同一记录。

来源Zenodo 记录 3832231 · canonical / citation_abstract_html_url

"https://zenodo.org/records/3832231"

开放状态

完全开放

Zenodo 标记为公开访问,TRAIN.zip 与 TEST.zip 均可直接下载,无需审批、注册或签署数据使用协议。

来源Zenodo API record 3832231 · metadata.access_right / files

"access_right": "open""; ""key": "TEST.zip""; ""key": "TRAIN.zip""

开放说明

公开内容是 TRAIN.zip 与 TEST.zip 两个压缩包。数据许可为 CC BY 4.0;配套的 DeepHistology 代码仓库使用 MIT 许可,只约束代码。

来源Zenodo 记录 3832231 / Zenodo API record 3832231 / GitHub API jnkather/DeepHistology · rights / access status / license

"The record and files are publicly accessible."; ""license": {"id": "cc-by-4.0"}"; ""name": "MIT License""

论文标题

Clinical-Grade Detection of Microsatellite Instability in Colorectal Tumors by Deep Learning

DOI、PubMed 与 PMC 记录的题名一致。

来源PubMed metadata / DOI metadata · TI / title

"Clinical-Grade Detection of Microsatellite Instability in Colorectal Tumors by Deep Learning."

下载链接

两个压缩包的直接地址:https://zenodo.org/api/records/3832231/files/TRAIN.zip/content 与 https://zenodo.org/api/records/3832231/files/TEST.zip/content。

来源Zenodo API record 3832231 · files / links.self_html

"links": {"self_html": "https://zenodo.org/records/3832231"}"; ""key": "TRAIN.zip""; ""key": "TEST.zip""

引用

@article{Echle_2020, title={Clinical-Grade Detection of Microsatellite Instability in Colorectal Tumors by Deep Learning}, volume={159}, ISSN={0016-5085}, url={http://dx.doi.org/10.1053/j.gastro.2020.06.021}, DOI={10.1053/j.gastro.2020.06.021}, number={4}, journal={Gastroenterology}, publisher={Elsevier BV}, author={Echle, Amelie and Grabsch, Heike Irmgard and Quirke, Philip and van den Brandt, Piet A. and West, Nicholas P. and Hutchins, Gordon G.A. and Heij, Lara R. and Tan, Xiuxiang and Richman, Susan D. and Krause, Jeremias and Alwers, Elizabeth and Jenniskens, Josien and Offermans, Kelly and Gray, Richard and Brenner, Hermann and Chang-Claude, Jenny and Trautwein, Christian and Pearson, Alexander T. and Boor, Peter and Luedde, Tom and Gaisa, Nadine Therese and Hoffmeister, Michael and Kather, Jakob Nikolas}, year={2020}, month=Oct, pages={1406–1416.e11} }

由 DOI 返回的正式 BibTeX。

来源DOI content negotiation · application/x-bibtex

"@article{Echle_2020, title={Clinical-Grade Detection of Microsatellite Instability in Colorectal Tumors by Deep Learning}"

数据许可

CC BY 4.0

来源Zenodo API record 3832231 · metadata.license

"license": {"id": "cc-by-4.0"}"

影响力

论文被引 337 次(Crossref)· DeepHistology 代码仓库 68 星 · 统计于 2026-06-27

来源Crossref Works API / GitHub repo page / GitHub repository API · message.is-referenced-by-count / stargazer button aria-label / stargazers_count

"is-referenced-by-count": 337; "68 users starred this repository"; ""stargazers_count": 68"

主要来源

数据集主页
https://zenodo.org/records/3832231(访问于 2026-05-29)
论文
Gastroenterology 159(4),2020 年 10 月 · doi:10.1053/j.gastro.2020.06.021 · PMID 32562722 · PMC7578071
托管记录
Zenodo 记录 3832231,版本 v1
元数据文件
Zenodo 记录元数据;TRAIN.zip 与 TEST.zip 文件清单;DeepHistology 说明文件(TCGA-CRC-DX 示例目录);论文补充表 S1(各队列临床病理特征)

补充表 S1 提供 TCGA 子队列的年龄、性别、部位、分期及 BRAF/KRAS 汇总;DeepHistology 说明文件只给出项目表格的结构。

来源Zenodo 记录 3832231 / PubMed metadata / PMC HTML 全文 / GitHub README / Supplementary Table S1 · Published May 18, 2020 | Version v1 / PMID metadata / PMCID block / TCGA-CRC-DX example / clinico-pathological features

"Published May 18, 2020 | Version v1"; "PMID- 32562722"; "PMCID: PMC7578071"; "An example for a possible data structure (project TCGA-CRC-DX) is"; "Table S1: Clinico-pathological features of each cohort."

临床与病理特征

器官

结直肠

补充表 S1 显示 TCGA 子队列同时包含结肠癌(321 例,75.4%)和直肠癌(105 例,24.6%)。

来源Zenodo 记录 3832231 / Supplementary Table S1 · description / cohort table

"These are histological images of colorectal cancer"; "colon cancer ... 321 (75.4%)"; "rectal cancer ... 105 (24.6%)"

肿瘤类型

结直肠腺癌

论文方法部分将 TCGA 子队列描述为结直肠腺癌患者;来源没有公开黏液腺癌等组织学亚型。MSIH / nonMSIH 是分子状态,不属于组织学分型。

来源PMC HTML 全文 · Materials and methods

"We retrospectively collected anonymized H&E stained tissue slides of colorectal adenocarcinoma patients"

主要分类字段

官方主任务
基于组织学图像块的 MSI 状态分类
MSI_Status
MSIH / nonMSIH;每个图像块继承所属患者的 MSI 状态

Zenodo 描述中阴性类写作 NonMSIH,压缩包目录名为 nonMSIH;本报告采用目录名。

来源Zenodo 记录 3832231 / TRAIN.zip 与 TEST.zip 官方 ZIP central directory · description / archive entries

"patients with MSI-H = MSIH; patients with MSI-L and MSS = NonMSIH"; "TRAIN/MSIH"; "TRAIN/nonMSIH"; "TEST/MSIH"; "TEST/nonMSIH"

临床信息

部分可用
字段层级取值
MSI 状态每名患者,继承到全部图像块MSIH / nonMSIH
UICC 分期队列汇总I 期 67(15.7%)· II 期 154(36.2%)· III 期 133(31.2%)· IV 期 59(13.8%)
BRAF队列汇总突变 56(13.1%)· 野生型 370(86.9%)
KRAS队列汇总突变 192(45.1%)· 野生型 234(54.9%)

只有 MSI 状态落到每个图像块;分期、BRAF、KRAS 只有队列汇总。DeepHistology 说明文件提到的 TCGA-CRC-DX_CLINI.xlsx 与 TCGA-CRC-DX_SLIDE.csv 只给出 PATIENT、FILENAME 等列名,没有公开逐例临床值。

来源Zenodo 记录 3832231 / GitHub README / Supplementary Table S1 · description / example data structure / clinico-pathological features

"all tiles inherited the label of the parent patient"; "CLINI table 'TCGA-CRC-DX_CLINI.xlsx'"; "SLIDE table 'TCGA-CRC-DX_SLIDE.csv'"; "mean age at Dx | 65.6"; "male | 211 (49.5%)"; "Region | US"; "BRAF mutant | 56 (13.1%)"; "KRAS mutant | 192 (45.1%)"

染色

H&E(苏木精-伊红)

论文与 Zenodo 描述均为常规 H&E 染色,未见免疫组化、免疫荧光等其他染色。

来源Zenodo 记录 3832231 / PMC HTML 全文 · description / Methods

"hematoxylin and eosin-stained slides"; "routine H&E histology"

是否罕见病

否

在 Orphanet 罕见病本体(ORDO 4.8)中检索 colorectal adenocarcinoma 与 colorectal cancer,均无精确匹配;近似结果是 Familial colorectal cancer Type X、Hereditary nonpolyposis colon cancer 等遗传性癌症易感综合征,与本数据集的结直肠腺癌范围不对应。

来源OLS4 ORDO ontology metadata / OLS4 ORDO search / OLS4 ORDO term detail · version / search / term detail

"version": "4.8""; ""numFound": 0"; ""label": "Familial colorectal cancer Type X""; ""label": "Hereditary nonpolyposis colon cancer""

罕见病名称

无

来源OLS4 ORDO ontology metadata / OLS4 ORDO search / OLS4 ORDO term detail · version / search / term detail

"version": "4.8""; ""numFound": 0"; ""label": "Familial colorectal cancer Type X""; ""label": "Hereditary nonpolyposis colon cancer""

人口统计

地区
美国
诊断时平均年龄
65.6 岁
性别
男 211(49.5%)· 女 213(50.0%)
原发部位
结肠 321(75.4%)· 直肠 105(24.6%)

均为补充表 S1 中 TCGA 子队列的汇总统计。来源没有公开种族或民族信息,也没有逐例人口统计文件。

来源Supplementary Table S1 · TCGA cohort statistics

"Region | US"; "mean age at Dx | 65.6"; "male | 211 (49.5%)"; "female | 213 (50.0%)"; "colon cancer | 321 (75.4%)"; "rectal cancer | 105 (24.6%)"

中心

多中心 · 来自 TCGA

论文称 TCGA 为多中心研究,患者主要来自美国;来源未列出具体供样医院。

来源PMC HTML 全文 · Materials and methods

"First, we used the publicly available Cancer Genome Atlas (TCGA, n=616 patients ...), a multicenter study with Stage I to IV patients mainly from the United States of America."

数据规模与格式

数据量

患者切片图像块
全部42342851,918
训练集 TRAIN28128419,557
测试集 TEST14214432,361
MSIH636415,002
nonMSIH36036436,916

按官方压缩包文件清单统计,训练集与测试集之间没有重叠患者。论文补充表 S4 记载 TCGA 队列为 426 名患者,比压缩包多 3 名;描述实际发布的数据时以文件清单为准。

来源TRAIN.zip / TEST.zip 官方 ZIP central directory / Supplementary Table S4 · archive entries / model table

"TRAIN/MSIH"; "TRAIN/nonMSIH"; "TEST/MSIH"; "TEST/nonMSIH"; "TCGA (N=426, 15% MSI)"

存储大小

约 3.36 GB:TRAIN.zip 1,279,103,781 字节,TEST.zip 2,084,051,578 字节

Zenodo 页面标注的 3.13 GB 是同一总量按 1024 进制换算的结果。没有单独的标注或元数据附件。

来源Zenodo 记录 3832231 / Zenodo API record 3832231 · schema.org / files

"contentSize": "3.13 GB""; ""key": "TRAIN.zip", "size": 1279103781"; ""key": "TEST.zip", "size": 2084051578"

有效图像数

51,918 个图像块

可分析的单位是图像块,不是原始全切片;它们来自 428 张切片、423 名患者。

来源TRAIN.zip / TEST.zip 官方 ZIP central directory · archive entries

"TRAIN/MSIH"; "TRAIN/nonMSIH"; "TEST/MSIH"; "TEST/nonMSIH"

数据模态

组织形态图像块

公开内容只有组织病理图像块,MSI 标签体现在目录结构中;没有发布掩膜、ROI 多边形、病理报告或逐例临床表。

来源Zenodo 记录 3832231 / TRAIN.zip 与 TEST.zip 官方 ZIP central directory · description / archive entries

"histological images"; "all tiles inherited the label of the parent patient"; "TRAIN/MSIH"; "TEST/nonMSIH"

数据详情

目录结构为 TRAIN/MSIH、TRAIN/nonMSIH、TEST/MSIH、TEST/nonMSIH,文件名由 TCGA 切片标识和图像块坐标组成,格式为 .jpg。肿瘤区域先由观察者在切片上人工勾画,再切成边长 256 μm 的图像块,保存为 512 像素(0.5 μm/像素),并用 Macenko 方法做颜色归一化。标签规则:MSI-H 为 MSIH,MSI-L 与 MSS 为 nonMSIH。训练集经随机下采样以平衡类别,测试集未下采样。

来源Zenodo 记录 3832231 / PMC HTML 全文 / GitHub README / TRAIN.zip 与 TEST.zip 官方 ZIP central directory · description / Methods / TCGA-CRC-DX example / archive entries

"Tumor tissue was outlined manually"; "cut into tiles of 256 μm edge length, saved as 512 px images"; "All image tiles were color-normalized with the Macenko method"; "TRAIN/MSIH"; "TEST/nonMSIH"

图像格式

图像块 · .jpg · 512 像素 · 0.5 μm/像素

原始全切片为 SVS 格式(补充表 S1),公开的是由其裁出的 JPG 图像块;扫描倍率未注明。

来源PMC HTML 全文 / Supplementary Table S1 / TRAIN.zip 与 TEST.zip 官方 ZIP central directory · Methods / cohort table / archive entries

"saved at a resolution of 0.5 μm per pixel"; "512×512x input layer"; "WSI format ... SVS"; ".jpg"

标本制备

未说明

来源只说明每名患者至少有一张组织切片、覆盖 I–IV 期,没有注明 FFPE、冰冻、活检或手术切除。

来源PMC HTML 全文 / Supplementary Table S1 · Materials and methods / cohort table

"For each patient, at least one histological slide was available"; "Stage I"; "Stage II"; "Stage III"; "Stage IV"

扫描仪

未说明

补充表 S1 只给出原始全切片格式 SVS,没有公开扫描仪厂商、型号、倍率或原始分辨率。

来源Supplementary Table S1 · TCGA cohort table

"WSI format ... SVS"

空间组学分辨率

不适用

该资源不含空间组学数据。

来源Zenodo 记录 3832231 / PMC HTML 全文 · description / Methods

"histological images"; "H&E stained tissue slides"

标注、任务与质控

任务类型

分类 · 基于组织学图像块的 MSI 状态分类

输入为图像块,输出为 MSIH 或 nonMSIH。

来源Zenodo 记录 3832231 / PMC HTML 全文 · description / Methods

"sorted by MSI status"; "all tiles inherited the label of the parent patient"; "We trained a deep-learning detector to identify samples with MSI from these slides"

任务说明

论文用法:输入结直肠癌 H&E 全切片肿瘤区域的颜色归一化图像块(512 像素,0.5 μm/像素),输出 MSIH 或 nonMSIH;图像块的预测结果在患者层面取平均。

来源Zenodo 记录 3832231 / PMC HTML 全文 · description / Methods

"sorted by MSI status"; "all tiles inherited the label of the parent patient"; "Tile-level predictions were averaged on a patient level"

图像来源

基于已有数据 · TCGA 结直肠癌全切片 · 托管于 Zenodo

图像块从已有的 TCGA 全切片中裁出并做了颜色归一化,不是新采集的切片;原始 SVS 文件可在 GDC 门户获取。

来源Zenodo 记录 3832231 / PMC HTML 全文 · description / Materials and methods

"derived from the TCGA database"; "All images and data from the TCGA study are publicly available at https://portal.gdc.cancer.gov"

标注来源

基于已有数据 · TCGA 经遗传学检测确定的患者级 MSI 状态

没有新绘制的掩膜或逐图像块标注,每个图像块继承所属患者的 MSI 状态。

来源Zenodo 记录 3832231 / PMC HTML 全文 · description / Methods

"all tiles inherited the label of the parent patient"; "Specimens with MSI were identified by genetic analyses."

多染色对齐

不适用

只有 H&E 一种染色,不涉及跨染色配准。

来源Zenodo 记录 3832231 / PMC HTML 全文 · description / Methods

"H&E stained tissue slides"; "color-normalized"

质控

人工质控 · 排除原因:质量不足、技术问题、无肿瘤组织

所有切片经观察者逐张人工复核,并由病理专家监督,确认含肿瘤组织且达到诊断质量;缺少分子信息的病例也被排除。质控针对切片与肿瘤区域,不是逐图像块的自动质控。论文讨论提到,外部测试集中有一张切片因技术伪影而模糊,人工复核没有发现。

来源PMC HTML 全文 · Methods / Discussion

"All slides were individually, manually reviewed"; "ensure that tumor tissue was present on the slide and the slide had diagnostic quality"; "excluded due to insufficient quality, technical issues, absence of tumor tissue"; "a technical artifact had resulted in a blurred image"

备注

口径差异

两处口径不同:患者数(压缩包文件清单 423,论文补充表 S4 为 426);阴性类写法(Zenodo 描述为 NonMSIH,目录名为 nonMSIH)。患者按 2:1 分为训练集与测试集。

来源Supplementary Table S4 / Zenodo 记录 3832231 / TRAIN.zip 与 TEST.zip 官方 ZIP central directory · model table / description / archive entries

"TCGA (N=426, 15% MSI)"; "Patients were split into training and test set in a 2:1 ratio"; "TRAIN/nonMSIH"; "TEST/nonMSIH"

Resource report

CRC-MSI resource report

Official name TCGA-CRC-DX · 38 fields, each quoting its source · sources checked in May–June 2026

On Zenodo, CRC-MSI is formally named TCGA-CRC-DX (record 3832231, titled “Histological image tiles for TCGA-CRC-DX, color-normalized, sorted by MSI status, train/test split”). Tiles are cut from the tumor regions of TCGA colorectal cancer H&E whole-slide images for microsatellite instability (MSI) classification, released as two archives, TRAIN.zip and TEST.zip. Each tile covers 256 μm, is saved at 512 × 512 pixels (0.5 μm/pixel) and is color-normalized; every tile inherits its patient's MSI status. Counted from the archive file listings, there are 423 patients, 428 slides and 51,918 tiles.

Basic record

Dataset name

TCGA-CRC-DX

Both the Zenodo title and the README of the official DeepHistology repository use TCGA-CRC-DX. CRC-MSI is an alias used for the MSI task, not the official name.

SourceZenodo record 3832231 · title / page title

"Histological image tiles for TCGA-CRC-DX, color-normalized, sorted by MSI status, train/test split"

Resource type

Dataset

Zenodo registers it as a dataset; it has none of the submission portal, leaderboard, hidden test labels or evaluation protocol a challenge or benchmark would have.

SourceZenodo API record 3832231 · metadata.resource_type / status

"resource_type": {"title": "Dataset", "type": "dataset"}"; ""status": "published""

Release date

2020-05

The Zenodo release date, 2020-05-18, is used. The related paper was published in October 2020 and is not taken as the data release date.

SourceZenodo API record 3832231 · metadata.publication_date

"publication_date": "2020-05-18""

Dataset homepage

DOI 10.5281/zenodo.3832231 points to the same record.

SourceZenodo record 3832231 · canonical / citation_abstract_html_url

"https://zenodo.org/records/3832231"

Access status

Fully open

Zenodo marks it as open access; TRAIN.zip and TEST.zip can both be downloaded directly, without approval, registration or a data use agreement.

SourceZenodo API record 3832231 · metadata.access_right / files

"access_right": "open""; ""key": "TEST.zip""; ""key": "TRAIN.zip""

Access notes

The public content is the two archives TRAIN.zip and TEST.zip. The data license is CC BY 4.0; the accompanying DeepHistology code repository uses the MIT license, which covers only the code.

SourceZenodo record 3832231 / Zenodo API record 3832231 / GitHub API jnkather/DeepHistology · rights / access status / license

"The record and files are publicly accessible."; ""license": {"id": "cc-by-4.0"}"; ""name": "MIT License""

Paper title

Clinical-Grade Detection of Microsatellite Instability in Colorectal Tumors by Deep Learning

The titles recorded by the DOI, PubMed and PMC agree.

SourcePubMed metadata / DOI metadata · TI / title

"Clinical-Grade Detection of Microsatellite Instability in Colorectal Tumors by Deep Learning."

Download link

Direct addresses of the two archives: https://zenodo.org/api/records/3832231/files/TRAIN.zip/content and https://zenodo.org/api/records/3832231/files/TEST.zip/content.

SourceZenodo API record 3832231 · files / links.self_html

"links": {"self_html": "https://zenodo.org/records/3832231"}"; ""key": "TRAIN.zip""; ""key": "TEST.zip""

Citation

@article{Echle_2020, title={Clinical-Grade Detection of Microsatellite Instability in Colorectal Tumors by Deep Learning}, volume={159}, ISSN={0016-5085}, url={http://dx.doi.org/10.1053/j.gastro.2020.06.021}, DOI={10.1053/j.gastro.2020.06.021}, number={4}, journal={Gastroenterology}, publisher={Elsevier BV}, author={Echle, Amelie and Grabsch, Heike Irmgard and Quirke, Philip and van den Brandt, Piet A. and West, Nicholas P. and Hutchins, Gordon G.A. and Heij, Lara R. and Tan, Xiuxiang and Richman, Susan D. and Krause, Jeremias and Alwers, Elizabeth and Jenniskens, Josien and Offermans, Kelly and Gray, Richard and Brenner, Hermann and Chang-Claude, Jenny and Trautwein, Christian and Pearson, Alexander T. and Boor, Peter and Luedde, Tom and Gaisa, Nadine Therese and Hoffmeister, Michael and Kather, Jakob Nikolas}, year={2020}, month=Oct, pages={1406–1416.e11} }

The formal BibTeX returned by the DOI.

SourceDOI content negotiation · application/x-bibtex

"@article{Echle_2020, title={Clinical-Grade Detection of Microsatellite Instability in Colorectal Tumors by Deep Learning}"

Data license

CC BY 4.0

SourceZenodo API record 3832231 · metadata.license

"license": {"id": "cc-by-4.0"}"

Impact

Paper cited 337 times (Crossref) · DeepHistology repository 68 stars · counted on 2026-06-27

SourceCrossref Works API / GitHub repo page / GitHub repository API · message.is-referenced-by-count / stargazer button aria-label / stargazers_count

"is-referenced-by-count": 337; "68 users starred this repository"; ""stargazers_count": 68"

Main sources

Dataset homepage
https://zenodo.org/records/3832231 (accessed 2026-05-29)
Paper
Gastroenterology 159(4), October 2020 · doi:10.1053/j.gastro.2020.06.021 · PMID 32562722 · PMC7578071
Hosting record
Zenodo record 3832231, version v1
Metadata files
Zenodo record metadata; TRAIN.zip and TEST.zip file listings; DeepHistology README (TCGA-CRC-DX example directory); the paper's Supplementary Table S1 (clinico-pathological features of each cohort)

Supplementary Table S1 gives cohort summaries of age, sex, site, stage and BRAF/KRAS for the TCGA subcohort; the DeepHistology README only shows the structure of the project tables.

SourceZenodo record 3832231 / PubMed metadata / PMC HTML full text / GitHub README / Supplementary Table S1 · Published May 18, 2020 | Version v1 / PMID metadata / PMCID block / TCGA-CRC-DX example / clinico-pathological features

"Published May 18, 2020 | Version v1"; "PMID- 32562722"; "PMCID: PMC7578071"; "An example for a possible data structure (project TCGA-CRC-DX) is"; "Table S1: Clinico-pathological features of each cohort."

Clinical and pathological features

Organ

Colorectum

Supplementary Table S1 shows the TCGA subcohort includes both colon cancer (321, 75.4%) and rectal cancer (105, 24.6%).

SourceZenodo record 3832231 / Supplementary Table S1 · description / cohort table

"These are histological images of colorectal cancer"; "colon cancer ... 321 (75.4%)"; "rectal cancer ... 105 (24.6%)"

Tumor type

Colorectal adenocarcinoma

The paper's methods describe the TCGA subcohort as colorectal adenocarcinoma patients; the sources give no histological subtypes such as mucinous adenocarcinoma. MSIH / nonMSIH is a molecular status, not a histological subtype.

SourcePMC HTML full text · Materials and methods

"We retrospectively collected anonymized H&E stained tissue slides of colorectal adenocarcinoma patients"

Main classification fields

Official main task
MSI status classification from histology tiles
MSI_Status
MSIH / nonMSIH; each tile inherits its patient's MSI status

The Zenodo description writes the negative class as NonMSIH, while the archive folder is named nonMSIH; this report uses the folder name.

SourceZenodo record 3832231 / official ZIP central directory of TRAIN.zip and TEST.zip · description / archive entries

"patients with MSI-H = MSIH; patients with MSI-L and MSS = NonMSIH"; "TRAIN/MSIH"; "TRAIN/nonMSIH"; "TEST/MSIH"; "TEST/nonMSIH"

Clinical information

Partly available
FieldLevelValues
MSI statusPer patient, inherited by all its tilesMSIH / nonMSIH
UICC stageCohort summaryStage I 67 (15.7%) · stage II 154 (36.2%) · stage III 133 (31.2%) · stage IV 59 (13.8%)
BRAFCohort summaryMutant 56 (13.1%) · wild type 370 (86.9%)
KRASCohort summaryMutant 192 (45.1%) · wild type 234 (54.9%)

Only MSI status reaches each tile; stage, BRAF and KRAS exist only as cohort summaries. The TCGA-CRC-DX_CLINI.xlsx and TCGA-CRC-DX_SLIDE.csv mentioned in the DeepHistology README only show column names such as PATIENT and FILENAME; no per-patient clinical values are public.

SourceZenodo record 3832231 / GitHub README / Supplementary Table S1 · description / example data structure / clinico-pathological features

"all tiles inherited the label of the parent patient"; "CLINI table 'TCGA-CRC-DX_CLINI.xlsx'"; "SLIDE table 'TCGA-CRC-DX_SLIDE.csv'"; "mean age at Dx | 65.6"; "male | 211 (49.5%)"; "Region | US"; "BRAF mutant | 56 (13.1%)"; "KRAS mutant | 192 (45.1%)"

Stain

H&E (hematoxylin and eosin)

Both the paper and the Zenodo description state routine H&E staining; no immunohistochemistry, immunofluorescence or other stains appear.

SourceZenodo record 3832231 / PMC HTML full text · description / Methods

"hematoxylin and eosin-stained slides"; "routine H&E histology"

Rare disease

No

Searching the Orphanet Rare Disease Ontology (ORDO 4.8) for colorectal adenocarcinoma and colorectal cancer gives no exact match; the nearest results are hereditary cancer-predisposition syndromes such as Familial colorectal cancer Type X and Hereditary nonpolyposis colon cancer, which do not correspond to this dataset's colorectal adenocarcinoma.

SourceOLS4 ORDO ontology metadata / OLS4 ORDO search / OLS4 ORDO term detail · version / search / term detail

"version": "4.8""; ""numFound": 0"; ""label": "Familial colorectal cancer Type X""; ""label": "Hereditary nonpolyposis colon cancer""

Rare disease name

None

SourceOLS4 ORDO ontology metadata / OLS4 ORDO search / OLS4 ORDO term detail · version / search / term detail

"version": "4.8""; ""numFound": 0"; ""label": "Familial colorectal cancer Type X""; ""label": "Hereditary nonpolyposis colon cancer""

Demographics

Region
United States
Mean age at diagnosis
65.6 years
Sex
Male 211 (49.5%) · female 213 (50.0%)
Primary site
Colon 321 (75.4%) · rectum 105 (24.6%)

All are cohort summaries for the TCGA subcohort from Supplementary Table S1. The sources give no race or ethnicity information and no per-patient demographics file.

SourceSupplementary Table S1 · TCGA cohort statistics

"Region | US"; "mean age at Dx | 65.6"; "male | 211 (49.5%)"; "female | 213 (50.0%)"; "colon cancer | 321 (75.4%)"; "rectal cancer | 105 (24.6%)"

Centers

Multicenter · from TCGA

The paper describes TCGA as a multicenter study with patients mainly from the United States; the sources do not list the contributing hospitals.

SourcePMC HTML full text · Materials and methods

"First, we used the publicly available Cancer Genome Atlas (TCGA, n=616 patients ...), a multicenter study with Stage I to IV patients mainly from the United States of America."

Data size and format

Data volume

PatientsSlidesTiles
All42342851,918
Training set TRAIN28128419,557
Test set TEST14214432,361
MSIH636415,002
nonMSIH36036436,916

Counted from the official archive file listings, no patient appears in both the training and test sets. The paper's Supplementary Table S4 records 426 patients for the TCGA cohort, 3 more than the archives; the file listings are used to describe the data actually released.

Sourceofficial ZIP central directory of TRAIN.zip / TEST.zip / Supplementary Table S4 · archive entries / model table

"TRAIN/MSIH"; "TRAIN/nonMSIH"; "TEST/MSIH"; "TEST/nonMSIH"; "TCGA (N=426, 15% MSI)"

Storage size

About 3.36 GB: TRAIN.zip 1,279,103,781 bytes, TEST.zip 2,084,051,578 bytes

The 3.13 GB shown on the Zenodo page is the same total converted in base 1024. There is no separate annotation or metadata attachment.

SourceZenodo record 3832231 / Zenodo API record 3832231 · schema.org / files

"contentSize": "3.13 GB""; ""key": "TRAIN.zip", "size": 1279103781"; ""key": "TEST.zip", "size": 2084051578"

Usable images

51,918 tiles

The unit of analysis is the tile, not the original whole-slide image; the tiles come from 428 slides of 423 patients.

Sourceofficial ZIP central directory of TRAIN.zip / TEST.zip · archive entries

"TRAIN/MSIH"; "TRAIN/nonMSIH"; "TEST/MSIH"; "TEST/nonMSIH"

Data modality

Tissue morphology tiles

The public content is only histopathology tiles, with the MSI label carried by the folder structure; no masks, ROI polygons, pathology reports or per-patient clinical tables are released.

SourceZenodo record 3832231 / official ZIP central directory of TRAIN.zip and TEST.zip · description / archive entries

"histological images"; "all tiles inherited the label of the parent patient"; "TRAIN/MSIH"; "TEST/nonMSIH"

Data details

Folders are TRAIN/MSIH, TRAIN/nonMSIH, TEST/MSIH and TEST/nonMSIH; file names combine the TCGA slide ID and the tile coordinates, in .jpg format. Tumor regions were first outlined manually on the slides, then cut into tiles of 256 μm edge length, saved at 512 pixels (0.5 μm/pixel) and color-normalized with the Macenko method. Label rule: MSI-H is MSIH; MSI-L and MSS are nonMSIH. The training set was randomly downsampled to balance the classes; the test set was not.

SourceZenodo record 3832231 / PMC HTML full text / GitHub README / official ZIP central directory of TRAIN.zip and TEST.zip · description / Methods / TCGA-CRC-DX example / archive entries

"Tumor tissue was outlined manually"; "cut into tiles of 256 μm edge length, saved as 512 px images"; "All image tiles were color-normalized with the Macenko method"; "TRAIN/MSIH"; "TEST/nonMSIH"

Image format

Tiles · .jpg · 512 pixels · 0.5 μm/pixel

The original whole-slide images are SVS (Supplementary Table S1); what is public are the JPG tiles cut from them. The scanning magnification is not stated.

SourcePMC HTML full text / Supplementary Table S1 / official ZIP central directory of TRAIN.zip and TEST.zip · Methods / cohort table / archive entries

"saved at a resolution of 0.5 μm per pixel"; "512×512x input layer"; "WSI format ... SVS"; ".jpg"

Specimen preparation

Not stated

The sources only say each patient has at least one tissue slide and that stages I–IV are covered; FFPE, frozen, biopsy or resection is not stated.

SourcePMC HTML full text / Supplementary Table S1 · Materials and methods / cohort table

"For each patient, at least one histological slide was available"; "Stage I"; "Stage II"; "Stage III"; "Stage IV"

Scanner

Not stated

Supplementary Table S1 gives only the original whole-slide format, SVS; the scanner maker, model, magnification and native resolution are not public.

SourceSupplementary Table S1 · TCGA cohort table

"WSI format ... SVS"

Spatial omics resolution

Not applicable

This resource has no spatial omics data.

SourceZenodo record 3832231 / PMC HTML full text · description / Methods

"histological images"; "H&E stained tissue slides"

Annotation, task and quality control

Task type

Classification · MSI status classification from histology tiles

The input is a tile; the output is MSIH or nonMSIH.

SourceZenodo record 3832231 / PMC HTML full text · description / Methods

"sorted by MSI status"; "all tiles inherited the label of the parent patient"; "We trained a deep-learning detector to identify samples with MSI from these slides"

Task description

As used in the paper: the input is a color-normalized tile (512 pixels, 0.5 μm/pixel) from the tumor region of a colorectal cancer H&E whole-slide image, the output is MSIH or nonMSIH, and tile predictions are averaged per patient.

SourceZenodo record 3832231 / PMC HTML full text · description / Methods

"sorted by MSI status"; "all tiles inherited the label of the parent patient"; "Tile-level predictions were averaged on a patient level"

Image origin

Derived from existing data · TCGA colorectal cancer whole-slide images · hosted on Zenodo

The tiles were cut from existing TCGA whole-slide images and color-normalized; they are not newly acquired slides. The original SVS files are available from the GDC portal.

SourceZenodo record 3832231 / PMC HTML full text · description / Materials and methods

"derived from the TCGA database"; "All images and data from the TCGA study are publicly available at https://portal.gdc.cancer.gov"

Annotation origin

Derived from existing data · patient-level MSI status determined by genetic testing in TCGA

No new masks or per-tile annotations were drawn; each tile inherits its patient's MSI status.

SourceZenodo record 3832231 / PMC HTML full text · description / Methods

"all tiles inherited the label of the parent patient"; "Specimens with MSI were identified by genetic analyses."

Multi-stain alignment

Not applicable

There is only one stain, H&E, so no cross-stain registration is involved.

SourceZenodo record 3832231 / PMC HTML full text · description / Methods

"H&E stained tissue slides"; "color-normalized"

Quality control

Manual QC · exclusion reasons: insufficient quality, technical issues, no tumor tissue

Every slide was reviewed manually by an observer under the supervision of a pathology expert to confirm it contained tumor tissue and was of diagnostic quality; cases without molecular information were also excluded. QC was applied to slides and tumor regions, not as automatic per-tile QC. The paper's discussion notes that one slide in an external test set was blurred by a technical artifact that the manual review missed.

SourcePMC HTML full text · Methods / Discussion

"All slides were individually, manually reviewed"; "ensure that tumor tissue was present on the slide and the slide had diagnostic quality"; "excluded due to insufficient quality, technical issues, absence of tumor tissue"; "a technical artifact had resulted in a blurred image"

Notes

Differences between sources

Two things differ: the patient count (423 in the archive file listings, 426 in the paper's Supplementary Table S4) and how the negative class is written (NonMSIH in the Zenodo description, nonMSIH as the folder name). Patients were split 2:1 into training and test sets.

SourceSupplementary Table S4 / Zenodo record 3832231 / official ZIP central directory of TRAIN.zip and TEST.zip · model table / description / archive entries

"TCGA (N=426, 15% MSI)"; "Patients were split into training and test set in a 2:1 ratio"; "TRAIN/nonMSIH"; "TEST/nonMSIH"