X-Message-Number: 28557
From: 
Date: Sat, 7 Oct 2006 01:40:31 EDT
Subject: Uploading technology (1.iv.3) MRI Brain Reader 4 

Uploading technology (1.iv.3) MRI Brain Reader 4

In the  preceeding message (#28547), I have argued for the use of electron 
resonance  imaging. If it is so good, why is it not in general use today ? I 

think the  problem is historical : MRI has been developed in the 70's-80's 
years. 
It was  difficult to get homogeneous magnetic fields in the sub milligauss 

range, so the  gradient field had to be in the gauss/inch to be both, manageable
and produce a  good picture definition. The continuous field could not then 
fall under one  kilogauss or so. One kgauss in electron magnetics resonance 

implies a frequency  near ten gigahertz, that was difficult for the electronics
amplifier.

One the other side, nuclear magnetics resonance allowed  the use of nearly 
off the shelf video amplifiers found in any TV set. So there  was no incentive 
to use ERI and there is no much today outside brain  reading.

Direct to home satellite TV technology has put on the  market amplifiers and 
components adapted to GHz amplifiers. There are off the  shelf amplifiers up 
to 45 GHz and ASIC circuits can be ordered up to 100 GHZ and  beyond. The 

technology used is the so called Pseudomorphic High Electron  Mobility 
Transistors 
(PHEMT).

Thre second problem, that is the  field homogeneity remains as before. A 
field can be produced with homogeneity of  one part per million in the volume 

looked for in MRI. If the main field is one  Tesla, that is 10,000 gauss, it can
be smoothed to .01 gauss. Because one gauss  translates into nearly ten MHz in 
electron magnetics resonance and the pixel  spectrum width is 1 kHz, the field 
sensitivity is .0001 gauss. So the pixels are  one hundred time smaller than 
the homogeneity scale of the magnetics field. If a  picture is taken in the 
X,Y plane, there will be many displacement in the third,  Z direction. A 3 

dimensional picture can't no more be built as a stack of flat  2D views. 30 
years 
ago, that was the end of the story.

Today, there  is one possibility around that : If the field roughness is 

stable, that is it  remains the same in time or evolves slowly in a predictable
way, the instrument  can be calibrated. For example, a set of straight fibers 

can be imaged, the  picture will look wavy because of the field inhomogeneities.
The "wavyness" is  then a mesure of the picture curvature or how it departs 
from flatness. Doing  that for any neighbor for each pixel give a map of the 
field distorting effect.  Taking then a 3-D picture of any object, the true 
position of each pixel can be  computed back. This process is called 

deconvolution, it is rather simple at the  elementary level, but ask for massive
computing 
power. This power is one the  market today, it was not 30 years ago. There are 
specialized circuits called  Digital Signal Processor (DSP) able to crunch up 
to 500 billions operations per  second on such a problem. Even for them, this 
is not a small  task.

Assume one pixel or voxel because this is 3D, is 10  nanometers on a side and 
the full picture is 10 cm (4") on a side. There will be  ten millions pixels 
in a row and 10^21 of them in the picture. If each needs 100  operations, the 
computing load is 10^23 or 100,000 billions of billions  operations. 200 DSP 
would take some 30 years to finish the job. The more  advanced DSP have phased 
local loop (PLL) problems i.e. synchronization  difficulties, beyond 230 

devices chained together, that is why I have assumed a  round number limit of 
two 
hundred of them. 

The full resolving  power is not yet asked for everywhere, it is usefull only 
at synaptic buttons,  there may be 10,000 of them per neuron and 10 billions 
of neurons. One synapse  can be defined with 100,000 voxels. That sum up to 
10^19 voxels, "only" one  percent of the preceeding brute force estimate. The 
computing time falls then to  4 months. What remains could be pictured at the 

micrometer scalle, giving an  economy of one million to one on the voxel number.
Even if this volume  include 99 percent of the full picture, it account only 
for 99/1,000,000 = .01%  of the computing load.

So, brain reading with ERI looks as a  possibility with current technology, 
even if the computing request is at the  front end of DSP capabilities.

Yvan Bozzonetti.






 Content-Type: text/html; charset="US-ASCII"

[ AUTOMATICALLY SKIPPING HTML ENCODING! ] 

Rate This Message: http://www.cryonet.org/cgi-bin/rate.cgi?msg=28557