fMRI Data Set for Visual Image Reconstruction


Here you can download our fMRI data set and MATLAB codes used to perform visual image reconstruction.

Summary

We release our fMRI data set used in Miyawaki Y et al. (2008): Visual image reconstruction from human brain activity using a combination of multiscale local image decoders. Neuron. Dec 10;60(5):915-29.

We used our custom-made machine learning algorithm “sparse multinomial logistic regression” to train the decoders of multiple scales, and combined them based on a simple linear model. This method works very well to reconstruct visual images from fMRI activity pattern with high accuracy, but there may be other methods that outperfoms our own. Please download the data set and test your idea on our data! The data set contains preprocessed fMRI data with the correspondig labels (stimulus). If you are familiar with machine learning but are new to fMRI, you are very welcome to download the data. Advanced knowledge about fMRI signals is not necessary.

The visual stimuli we used to train the reconstruction model are just random image patterns, which are unbiased toward any specific spatial patterns (see Figure 1 in the paper). In this sense, the experiment we perform is very similar to that of receptive field identification using white noise patterns. Beside visual image reconstruction, you may use the data for differnt purposes such as identification of receptive field of human visual cortex, investigation of spatial linearity of BOLD signals, etc. Those whos are not interested in visual image reconstruction but are interested in basic physiological questions are also very welcome to download the data, and test hypotheses on our data!

Currently, we have not released completely raw fMRI files. But we are also happy to release them in ANALYZE format upon your requests.

History

August, 7 (2009): released beta version of the data set.

May, 15 (2010): added MNI coordinate values for each voxel. Descriptions for the coordinate values are included in the manual pdf.

May, 31 (2012): released Matlab codes for visual image reconstruction.

Contents

The data set cosists of a single zip file (‘Data_Reconst_Code_public.zip’). Extracting the zip file, you will have the data file in Matlab format, Matlab codes for image reconstruction and the pdf document manual. A total of file size is about 900MB. The data set contains preprocessed fMRI signals (slice timing correction, realignment, coregistration, and reslice were applied) for the random image session, which is necessary to train the reconstrcution model, and for the figure image session, which is for illustrating the generalization performance of the image reconstruction model (see Figure 2 in the paper). The data set corresponds to subject S1 in the paper.

Environment

The data is stored by Matlab version 7.4 compatible format. Please use the Matlab later than that version. Brain Decoder Toolbox (BDTB) ver. 1.2.2 is required to perform Matlab codes for visual image reconstruction. BDTB is available from here.

Download

Go to Visual Image Reconstruction Project page in brainliner.jp. You can download “Data_Reconst_Code_public.zip” in “Supplementary Files”. All data and the MATLAB codes are included in the zip file.

Note

You can use the data freely. We do not require any authorships. Please cite the following our paper when you publish your results using our data.

References

Miyawaki Y, Uchida H, Yamashita O, Sato MA, Morito Y, Tanabe HC, Sadato N, Kamitani Y (2008). Visual image reconstruction from human brain activity using a combination of multiscale local image decoders. Neuron. Dec 10;60(5):915-29.

Contact

Any comments, questions, requests are welcome. Please contact the following address:

Department of Neuroinformatics
ATR Computational Neuroscience Laboratories
2-2-2 Hikaridai, Seika-cho, Soraku-gun, Kyoto 619-0288

E-mail: dni-info (at) atr.jp
http://www.atr.jp/index_j.html