Messidor is an image classification problem. The data consists of 1200 eye fundus images from 654 diabetes and 546 healthy patients. Each image from the original data is rescaled to 700×700 pixels and split up into patches of 135×135 pixels. Patches which do not have a sufficient amount of foreground are discarded. The features used are: intensity histogram of RGB channels for 26 bins, mean of local binary pattern histograms of 20×20 pixel grids, mean of SIFT descriptors, and box count for grid sizes 2 to 8. Some of the features return NaNs, replacing by zero is advised.

Original source

The original data is kindly provided by the Messidor program partners (see

This dataset has been represented as a MIL problem by Dr. Melih Kandemir.

When using the dataset, please cite the following papers:

title={Feedback on a publicly distributed image database: the {M}essidor database},
author={Decenci{\`e}re, Etienne and Zhang, Xiwei and Cazuguel, Guy and Lay, Bruno and Cochener, B{\'e}atrice and Trone, Caroline and Gain, Philippe and Ordonez, Richard and Massin, Pascale and Erginay, Ali and Charton, B{\'e}atrice and Klein, Jean-Claude},
journal={Image Analysis and Stereology},

title={Computer-aided diagnosis from weak supervision: A benchmarking study},
author={Kandemir, Melih and Hamprecht, Fred A},
journal={Computerized Medical Imaging and Graphics, in press},



Leave a Reply

Your email address will not be published. Required fields are marked *