Comments on life, science, business, philosophy, and religion from my personal public health viewpoint
Friday, November 02, 2007
Decentralized sense-making in a cluttered world
If central planning isn't helpful for sense-making in a complex and cluttered world, what is?
The world of "image-processing" in computing has come up with some techniques that seem interesting models for action. I want to describe one that I've used in the past. You don't need any math for this. I tried to make it easy to follow.
The problem we had involved finding the edges of a brain tumor on a 3-dimensional Magnetic Resonance Imaging image. This is actually a set of "slices", stacked like a deck of cards, across a section of the brain.
Each slice looks something like this picture, which is a cross-section image I pulled off the web from the NIH Image database of public sample images. That's a vertical "slice" through someone's head, facing to the left. (Note - the person isn't actually sliced or injured - the computer just makes it look that way!)
Maybe if you think of baking an orange into a loaf of bread, and then running it through a bread slicer -- you get the image of a stack of slices, starting with all bread and no orange, then really small circles of orange, then slices with larger circles, then smaller again, and finally bread slices with no orange at all. Our job is to find the orange in the pictures of the slices of bread and reconstruct what it looks like in 3-D.
If there is or might be a tumor, it's important to find the edges as accurately as possible, based on these kinds of images. That's not as easy as you might think, because when you zoom to the high magnification, the images are actually pretty blurry and "noisy" and hard to read as to where, exactly, an "edge" is.
Here's some structure in a brain, probably not a tumor, for illustration. If you click on the image, you can zoom it up and see some sort of black dot with a white border fairly easily in the upper right, "Slice #19". But if you look at the previous slice, the next "card in the deck of cards",
"Slice #18", the edges of this are less distinct and this slice of the orange is smaller.
Similarly around slice 20, maybe we can still be fairly sure we "see" the edges of the white structure, but by slice 21 it's not clear what's that structure, and what's just normal tissue.
And, we're using the magic of human eyes. We want some way the computer can do a better job than people at finding the edges of a structure, once a trained radiologist points it out. (This was all done over a decade ago and I suspect they have way better tools today, by the way.)
Anyway, let me describe how the edges can be found. Look at it on one slice first. Imagine surrounding the tumor with a line of people attached by stretchable elastic cords or "slinkies" or springs. In this picture I just drew a red dot instead of person, but you get the idea. Pretend that's the view from above of many people with red hats connected to each other with adjustable bungee cords.
Then, you ask each person, when he gets to a place where he looks down and sees dark changing to light rapidly, that might be the edge of the tumor, so he should dig in his heels and try to stay there. But, at the same time, you start making the springs stronger, pulling people towards each other.
As you do that, initially with the springs fairly stretchy and loose, the circle starts being pulled smaller and smaller, like a drawstring tightening on a purse. When each person gets to what seems like it might be an edge, they try to drag their heels and stop moving, independently.
After this has gone on a while, you may end up with something like this:
You can see that most of the people have found the edge of the tumor and dug in their heels there. But people #1 and #2 found a bright edge that is probably just "noise". And the people numbered "3" have found something that it's hard to tell if that's tumor or noise.
How should they decide?
In this technique, if you just start tightening the springs, at some point the collective pulling force of the majority will break #1 and #2 loose from the feature they are snagged on, and they'll snap into place around the tumor.
Based on just this slice, the people labelled 3 may not move, because maybe that's actually an edge. (Tumors don't have to be round - they can be irregular.)
Well, how do the people at #3 decide? Here's the trick. While all this is going on on this slice,
the same thing is going on on all the other slices, and springs connect the dots / people across the slices. In other words, we actually have a sphere of dots / people, connected by springs, kind of like an over-inflated balloon, and we let it slowly deflate in three dimensions at once,
around the feature in all the slices.
In other words, there is not enough information in the vicinity of any one person to be able to sort out image from noise with certainty. But, most of them are nearly right. We just don't know which ones that is. So, within each slice (or region) the people consult with each other, while at the same time they are consulting across regions as well, with a mix of believing their own eyes, and enough humility to know, at some point, to let go and go with the crowd.
This seems like a very simple plan, and no computer is required. It turns out to be a very powerful technique ("algorithm") that does a remarkably good job at sorting out "noise" from "signal" in 3-dimensions, with only trivial programming required.
Each person / dot simply has to pay attention to what it sees ("independent investigation"), but balance that with consulting with neighboring people and at some point yielding to peer pressure and moving into line. If the balance of these two competing forces is right, the overall network turns our to be a very powerful analog computer that can solve a problem we have trouble even defining well.
No single person ever needs to "see" everything or see "the big picture" - he just needs to see his neighbors, compare notes, argue for his position, and, if it seems warranted, yield to the majority. If enough different dots do this, coming in from enough different directions at once ("diversity"), and remain independent and yet consulting ("unity in diversity") the algorithm works. The powerful solution "emerges" from each person's behavior.
In image processing this is called an "adaptive contour" technique. It is part of the larger class of techniques called "swarm computing" that is becoming increasingly popular as "the power of crowds" is increasingly being appreciated.
An area this could be used is in any sort of boundary measurement, or in aligning fragments of images to make a coherent overall picture. Examples of these, and my US Patent 5613013 in image alignment using effectively a swarm technique, are described on my web site here.
Wade
Subscribe to:
Post Comments (Atom)
1 comment:
Cool! Now we just need to get those people in line... Is there a technique for herding cats?
Post a Comment