Dataset

A metaphase chromosome image dataset with karyotype boxes (Tseng et al. 2023, G-banded)

6,993 G-banded metaphase cell images with 321,838 chromosome bounding boxes, in a 24-class karyotype set and a single-class set. From Tseng et al., Scientific Data 2023, CC BY 3.0. Browse it in DataTorch.

Contents

  1. What it is
  2. Browse it in DataTorch
  3. How to split it
  4. What people use it for
  5. Source and citation
Contents

What it is

Eight metaphase spreads with a box drawn around every chromosome. Six show coloured boxes named by chromosome class, two show boxes of one class.
Eight images from datatorch/metaphase-chromosomes-cil54816. Six from the 24-class set, with each box named by its chromosome class, and two from the single-class set.

To read a karyotype, a cytogeneticist photographs a cell caught in metaphase, when its chromosomes are condensed and spread apart, then finds each chromosome and decides which of the 24 it is. A model that helps with that has two jobs: find every chromosome in the spread, and name it. This dataset has labels for both.

The images come from the prenatal cytogenetics laboratory of Taichung Veterans General Hospital in Taiwan. The paper says they were compiled from the chromosome studies of 1,598 fetuses between 2014 and 2021, with the approval of the hospital's review board and the consent of the subjects. The Cell Image Library record describes amniotic cells stained with trypsin and Wright stain for G-banding, photographed through the microscope.

FactValue
PaperTseng et al., An Open Dataset of Annotated Metaphase Cell Images for Chromosome Identification, Scientific Data 10, 104 (2023)
RecordCell Image Library CIL:54816, doi 10.7295/W9CIL54816
Images5,000 labelled with 24 chromosome classes and 2,000 labelled with a single chromosome class, as the paper describes them
Chromosomes229,852 in the 24-class set, per the paper
Labelsbounding boxes in Pascal VOC XML, one file per image
StainG-banding (trypsin and Wright stain)
Sizeone zip, 1.21 GB
LicenseCC BY 3.0

The values in this table are the paper's and the record's. The counts of the hosted copy are below and differ slightly, for reasons given there.

The labels are boxes, not outlines. If you need the shape of each chromosome, for straightening or for separating chromosomes that touch, this dataset does not have it.

Browse it in DataTorch

Dataset: datatorch/metaphase-chromosomes-cil54816, 7,007 files

Annotated metaphase chromosome images (Tseng et al. 2023)

Annotated metaphase chromosome images from Tseng et al. 2023 (Cell Image Library CIL:54816): clinical prenatal G-banded metaphase cell images from the cytogenetics laboratory of Taichung Veterans General Hospital, collected with review-board approval and patient consent as the paper states. 6,993 images with bounding boxes (boxes, not masks) in two sets: 4,997 images with 229,714 boxes in 24 chromosome classes (A1-G22, X, Y) and 1,996 images with 92,124 boxes in a single class (chromosomes); 1,843 images appear in both sets. Changes from the source: files renamed to sequential ids, the path field removed from the annotation files, label tables and split lists re-keyed to the new names, image metadata removed, model weight files not included, and five images left out because a lab label was readable in the image. A small share of the remaining images may carry a lab label in the pixels, as published by the source. License: CC BY 3.0. Source: http://www.cellimagelibrary.org/images/54816. Please cite: Tseng JJ, Lu CH, Li JZ, Lai HY, Chen MH, Cheng FY, Kuo CE. An Open Dataset of Annotated Metaphase Cell Images for Chromosome Identification. Sci Data 10, 104 (2023). doi:10.1038/s41597-023-02003-7. Dataset: CIL:54816, doi:10.7295/W9CIL54816.

Open →

The public metaphase chromosome project in DataTorch, showing 7,007 files and thumbnails of metaphase spreads
The hosted copy, datatorch/metaphase-chromosomes-cil54816, on November 18, 2026

The hosted copy holds 6,993 images with 321,838 bounding boxes in 25 labels: the 24 chromosome classes and one label, chromosomes, for the single-class set. The two sets sit in two folders.

SetFolderImagesBoxesLabels
24-classimages/boxes24/4,997229,714A1 to G22, X, Y
Single-classimages/single/1,99692,124chromosomes

Open any image to see its boxes, filter by label to look at one chromosome class across cells, and correct a box where you disagree with it. An image carries 43 to 57 boxes, 46 at the median, which is what a human cell should show. The images are small JPEGs, between 360 and 1,360 pixels wide.

The copy is not byte-for-byte the archive. The project's notes list every change, and these are the ones that matter:

  • Files are renamed. The source names its files by hospital case number. Every image here has a shuffled id instead (boxes24_000101.jpg, single_000050.jpg), the label tables and split lists were rewritten to the new ids, and the mapping back is not published.
  • Image metadata is removed. The pixels are unchanged.
  • Five images are left out, because a hospital case code was readable in the picture. That is why the 24-class set has 4,997 of the paper's 5,000 images and the single-class set 1,996 of 2,000, and it is consistent with the hosted 24-class set holding 138 fewer boxes than the paper's 229,852.
  • A small share of the remaining images may carry a lab label in the pixels, as the source published them. They were not searched or edited.
  • 195 boxes were clipped to the image edge, 77 in the 24-class set and 118 in the single-class set, where the source box ran past it.
  • The model weight files in the archive are not included.

The project also keeps the authors' tables, re-keyed to the new ids: the karyotype of each image in three files (normal, number abnormalities, structural abnormalities), the list of images the authors rated difficult, and the original-style XML.

How to split it

Use the authors' lists. The project keeps them as tables/train.txt and tables/test.txt, which its notes give as 3,742 and 1,249 rows.

Two things to watch if you make your own split:

  • The two sets overlap. 1,843 images are in both sets as identical files under two ids, and the tables pair 1,833 of them. Remove the pairs before you pool the sets or test one against the other.
  • Several images come from one fetus. About 5,000 images from 1,598 fetuses means a random split by image will often put cells of the same case on both sides. The hosted copy cannot group by case, because the case numbers were removed, and we do not know whether the authors' lists are grouped that way.

What people use it for

Training and comparing chromosome detectors and classifiers on real clinical spreads, where chromosomes bend, touch and cross. The single-class set trains the finding step on its own, and the 24-class set the naming. The karyotype tables let you look at how a model does on cells with a number or structural abnormality, which is where a classifier that has learned "two of each" goes wrong.

It is a research dataset. Nothing here is a diagnostic tool, and a model trained on one laboratory's stain and camera should be expected to do worse on another's.

Source and citation

Source: http://www.cellimagelibrary.org/images/54816 (read 2026-11-18). Cite the paper: Tseng JJ, Lu CH, Li JZ, Lai HY, Chen MH, Cheng FY, Kuo CE. An Open Dataset of Annotated Metaphase Cell Images for Chromosome Identification. Scientific Data 10, 104 (2023). https://doi.org/10.1038/s41597-023-02003-7. And the dataset: Lu CH, Kuo CE, Tseng JJ. CIL:54816. Cell Image Library (2022). https://doi.org/10.7295/W9CIL54816. The license is CC BY 3.0, which asks that changes be indicated. The changes in the hosted copy are listed above and in the project.

Label your own metaphase images

Start a project that already has these chromosome labels, add your own images and draw the boxes with your team in the browser.

Frequently asked questions

Is there an open dataset of metaphase chromosome images with labels?

Yes. Tseng et al. 2023 is distributed by the Cell Image Library as CIL:54816 under CC BY 3.0. The paper describes 5,000 images labelled with 24 chromosome classes and 2,000 labelled with a single chromosome class, from 1,598 fetuses.

Does it have segmentation masks of the chromosomes?

No. Both sets are bounding boxes in Pascal VOC format. The second set is sometimes described as segmentation, but every annotation file in it is a box file. It suits detection and classification, and it does not give you chromosome outlines.

What are the 24 classes?

The 22 autosomes plus X and Y. In the hosted copy each autosome label carries its Denver group letter and its number, from A1 to G22.

Can I combine the two sets for training?

Carefully. 1,843 images are in both sets as identical files under two ids. If you train on one set and test on the other, or pool them and split at random, the same image lands on both sides. The hosted tables pair most of the duplicates so you can remove them.

What is the license?

Creative Commons Attribution 3.0 (CC BY 3.0), stated on the Cell Image Library record. Cite the paper, doi 10.1038/s41597-023-02003-7, and the dataset, doi 10.7295/W9CIL54816.

Get the data

Open the hosted copy on DataTorch.

Browse the files in your browser, then download the dataset with a free account.

Open the datasetStart free

Related

Tools and community for image annotation: label, review, and share datasets, from everyday photos to the imagery only specialists can read.

Platform

PlatformImaging labsExplore datasetsEnterprisePricing

© 2026 DataTorch. All rights reserved.