MHIST Histopathology Image Dataset

MHIST stands for Minimalist Histopathology image analysis dataset, which is a binary classification image dataset. The dataset contains 3,152 fixed-size colorectal polyp images, and the gold-standard annotation of each image is determined by 7 gastrointestinal pathologists by majority vote. MHIST also includes the degree of agreement of the annotators for each image. As a minimal dataset, the dataset takes up less than 400MB of disk space. The ResNet-18 baseline can be trained to converge on MHIST in just 6 minutes using 3.5 GB of memory on an NVIDIA RTX 3090.

As a use case, the researchers used MHIST to study natural problems that arise in histopathology image classification, such as how dataset size, network depth, transfer learning, and high-divergence examples affect model performance.