MedMNIST

MedMNIST v2 is a collection of standardized biomedical image datasets. MedMNIST loads one of them as images (N, 3, size, size) uint8 and labels (N,).

Yang et al., MedMNIST v2: A large-scale lightweight benchmark for 2D and 3D biomedical image classification, Scientific Data 2023. Paper · Data · CC BY 4.0

Usage

from torch_world.datasets import MedMNIST

blood = MedMNIST("bloodmnist", split="train", size=64)
item = blood[0]               # {"image": (3, 64, 64) uint8, "label": ()}

blood.images, blood.labels    # (N, 3, 64, 64), (N,)
blood.modality                # "blood smear microscopy"
blood.classes                 # ["basophil", "eosinophil", ...]
blood.malignant               # class ids that are malignant (empty here)
blood.info                    # "bloodmnist (train): blood smear microscopy, 8 classes, 11959 images"

argument
name	one of the datasets below
split	"train", "val" or "test"
size	28, 64, 128 or 224
root	cache directory, ~/.cache/torch_world/medmnist by default
download	fetch the file if it is missing; otherwise raise FileNotFoundError

The .npz file is downloaded once from Zenodo and checked against its MD5 on every load. It is read without NumPy. Grayscale images are repeated to three channels, so every dataset feeds the same encoder.

Available datasets

name	modality	classes	malignant classes	train / val / test
bloodmnist	blood smear microscopy	8 cell types	none	11,959 / 1,712 / 3,421
breastmnist	breast ultrasound	malignant, benign lesion	malignant lesion	546 / 78 / 156
dermamnist	dermatoscopy	7 skin lesions	actinic keratosis, basal cell carcinoma, melanoma	7,007 / 1,003 / 2,005
organamnist	abdominal CT	11 organs	none	34,561 / 6,491 / 17,778
pathmnist	colon pathology	9 tissue types	colorectal adenocarcinoma epithelium	89,996 / 10,004 / 7,180

Split sizes are those published with MedMNIST v2.

Reference

MedMNIST in the API reference.