Since I’ve decided to return to school to get my PhD, I figured it would be a good idea to start reviewing research papers again. I must admit that its been a while since I’ve reviewed a research paper and I’ve never reviewed a neuroscience paper before. None-the-less, here is my attempt at a review.

Summary: The paper presents research using fMRI analysis of subjects’ brain patterns when viewing various 10×10 block images. The method uses multiple points of analysis of the brain and the image reconstruction algorithm requires training to select voxel weights.

They present analysis of random image identification using a subset of all possible random images and show the rate of successful identification declines as the size of the potential image set increases. The authors explain their voxel selection method and perform tests to determine the quantifiable benefit of their voxel selections over other methods. Next they explore their image reconstruction technique to determine how well their multiscale method performed over single scale methods. My guess is the multiscale method was an after-thought once they reviewed the performance of the single scale data. The multiscale image did perform better at correctly matching the presenting image, however, I didn’t understand how they were actually calculating ‘performance’ or ‘eccentricity’ despite relying on it in the analysis. It could be common knowledge in the field that I’m just not familiar with yet.

The authors do an excellent job at reviewing their own work in the discussion section and present a handful of potential future areas of research. They finish with a detailed explanation of their experimental procedures which I found particularly interesting because it covered details of their data gathering, filtering, and processing using different sources such as correlating retinotopy mapping to the fMRI data (though I didn’t understand it completely). It gave me a good insight into the experimentation process used in neuroscience.

Questions: I’m not sure if its just my lack of knowledge about the type of data the fMRI and retinotopy mapping produces or if the authors were just vague but their observations of the “Weight Distribution on the Cortical Surface” seemed a bit sparse. It didn’t seem to suggest why the weights one way or another were more or less beneficial to their experimental results. Perhaps this is an area for future research or perhaps my knowledge in the field is showing itself. They did follow that section up with experimenting with different methods and it might seem more obvious why if I knew a bit more about retinotopy.

Future Work: It would be interesting to see if similar techniques could be used to identify the main component of a larger picture. My thinking is that if they can reconstruct a 10×10 image what ability could the same or an adapted technique have at reconstructing a 10×10 piece of a bigger picture? For instance a picture in which a contrasting foreground image of 10×10 was placed against a larger random background. What happens when the test data is three dimensional instead of a simple 2d image?

Could a similar technique be used to identify simple shapes in more complex scenes? Would other areas of the brain come into play when using actual objects or more complex scenes? How well does this technique scale up beyond the 10×10 images? How can we identify color?

It would be interesting to see the effectiveness of various types of multivoxel pattern decoder methods in relation to to same experiment or even with scaling it up. The selected technique worked for this instance of 10×10 block images but will it still work if the image becomes more complex like 100×100 or even 1000×1000? What is the effectiveness of different methods as the image complexity increases? Perhaps other reconstruction methods work better for more complex images. I’m curious what the most complex reconstruction is to date and what methods were used.

One of their conclusions is that the multiscale image combination method contributed to the positive results. It seems highly coupled with the voxel selection technique. It would be interesting to explore other image combination techniques in relation to different voxel selection methods. Perhaps another voxel selection / image combination technique would produce more accurate results.

Conclusions, Interests and Reflections:
The authors mention they used a statistical learning algorithm called “sparse logistic regression” to train the weights of voxels. I haven’t come across this learning algorithm before and am interested in learning more about it.

It has become clear to me that I know far too little about brain imaging technology to understand this paper completely. I intend to study different brain imaging technologies like fMRI next so that I can better understand some of the details in future papers.

Overall I think this was an extremely well researched paper. The results were clearly presented and the authors seemed to give a thorough analysis of their methodology.

Visual Image Reconstruction from Human Brain Activity using a Combination of Multiscale Local Image Decoders – Yoichi Miyawaki, Hajime Uchida, Okito Yamashita, Masa-aki Sato, Yusuke Morito, Hiroki C. Tanabe, Norihiro Sadato, and Yukiyasu Kamitan