Multi Cancer Dataset
130,000 uniform images spanning eight cancer types and twenty-six subclasses — assembled so that researchers can stop spending their first week on file plumbing.
Images
130,000 · JPEG · 512×512
Taxonomy
8 classes · 26 subclasses
Licence
CC BY-NC-SA 4.0
Kaggle signal
Bronze medal · 141 notebooks
01 · What it contains
IndexChapter 01
What the dataset is
A consolidated corpus of cancer imagery: eight main cancer classes, each broken into subclasses, every image standardised to 512×512 JPEG with a predictable <subclass>_<serial>.jpg filename. Each subclass folder holds 5,000 images, so the classes are balanced by construction.
Chapter 02
Why it exists
The research it supports needed one model to handle many cancers. The imagery for that existed — but scattered across half a dozen separate Kaggle and Figshare datasets, each with its own resolution, directory convention, and naming scheme. Before any modelling could begin, someone had to normalise them. That normalisation turned out to be the more reusable contribution.
Chapter 03
Structure, as published
Select a class to see its verified share of the corpus. The distribution below uses the counts from the Kaggle data card; the tiles are illustrative, not real medical scans.
Class distribution · 130,000 images · 8 classes
Selected class
Cervical
5 subclasses
25,000 images · 19.2% of corpus
Illustrative tiles — not real samples
Images were augmented with Keras’ ImageDataGenerator — modest rotation, shift, shear, zoom, horizontal flip, and brightness variation — then cropped and renamed for uniformity.
Chapter 04
How people use it
It is the most-used artifact I have published. As of July 2026, the Kaggle data card shows a bronze dataset medal, 163 upvotes, and 141 public notebooks built on it — training everything from VGG16 and MobileNetV3 baselines to lymphoma-specific classifiers.
Chapter 05
Responsible use
Cite this
Obuli Sai Naren. (2022). Multi Cancer Dataset [Data set]. Kaggle.
doi.org/10.34740/KAGGLE/DSV/3415848 ↗Lessons
What it taught me
- The unglamorous work — renaming, resizing, balancing — is what other researchers actually reuse. The paper got cited; the dataset got built upon.
- Provenance is a feature. Linking every source dataset made this trustworthy in a field where mystery corpora are common.
- A good README is an API. The dataset scores highly on usability because the structure is documented, not because the images are special.