Research Interests: Past & Present
This is an overview of science & engineering topics of interest, past and present.
————————
Brain-Computer Interface
BCI: Research allowing direct communication between the computer and the human mind. A Brain-Computer interface could be used to control prosthetic devices, significantly enhancing the quality of life for handicapped individuals. The NIH has formed the Neural Prosthesis Program, which holds yearly workshops and funds some of the research.
BCI research is proceeding along invasive, intra-cortical lines as well as more data-processing intensive EEG-based approaches. The latter methods affix EEG leads on the scalp, record brain waves, and employ powerful computer methods to decipher the results. Noise is a problem, so researchers have embraced the more invasive approach of implanting chips directly into the brain.
Invasive Research
The Nicolelis lab at Duke and the Chapin group at SUNY-Brooklyn have showed amazing results with monkeys and rats.
Toby Howard, a researcher at the Univ. of Manchester, has a nice collection of links, although its a bit dated. The Lab of Brain-Computer Interfaces, Technical University of Graz, has an active group researching BCI, both through EEG and implanted electrodes (ECoG).
Cyberkinetics is a spin-off from Brown University research using implanted electrodes as the interface.
Neural Signals Inc is a spin-off from Georgia Tech research also using implanted electrodes, in this case small glass tubes with coatings that encourage neuronal growth.
The Dobelle Institute focuses on artificial vision systems for the blind. They moved to Portugal in order to proceed with human experiments, where processed video signals are fed to an intra-cortical implant.
NeuroPace develops implantable neurostimulators that preempt epileptic seizures. It’s technology might be used for a number of conditions.
Non-invasive Research
The principal researchers in this area include the German group in Graz and the Wolpaw group in Wadsworth, NY.
————————
Computer Optimization of Radiotherapy
At Varian Medical Systems, I researched ways for a computer to automatically generate radiotherapy plans that optimally satisfy a medical practitioner’s constraints. Given a target volume, many normal tissue volumes, and some number of radioactive seeds, the goal was to construct a computer program that automatically places seeds so that the radiation optimally covers the target volume while sparing as much of the normal tissue as possible. The problem was confounded by the phrase “optimally covers” since each physician has their own notion of optimal coverage.
The current version of the Varian system includes a rules-based dose-optimization module that resulted from this research. Most of this work is proprietary and held as trade secrets or patents.
————————
Neurosurgical Navigation
Although 3D visualization provides stunning pictures, does it really improve the care of patients? In many ways, the answer is not so obvious. Physicians have been trained to look at medical images in two-dimensions ever since the x-ray was used clinically. Radiologists, the specialists in medicine who look at images, have just learned the benefits of 3D visualization for diagnosis in the past decade. But most of the time, physicians can do the job of diagnosis using just regular 2D films. It doesn’t take Doc Genius to figure out the patient has a massive subdural hematoma when one 2D slice of a CT scan shows a huge dark area inside the skull.
So when does 3D visualization really pay off? Aside from subtle injuries, the third dimension is important when you start doing therapy or surgery. When attempting to remove or irradiate a “bad” area, you have to spare as much of the normal tissue as possible, and the only way to plan such operations is to have a precise and visually clear map of the entire volume.
Many of the problems encountered in neurosurgical navigation are also present when physicians try to treat patients with radiation.
————————
Digital Humans
In 1991, the National Library of Medicine began funding one of the most fascinating computer projects of the 20th century, the Visible Human Project. The project used two donated cadavers, one male and one female, to construct an extensive image database of human anatomy amounting to more than 50 gigabytes (or 50,000 megabytes) of data. The unprocessed data consists of over 6,000 color photographs of cross-sectional anatomy at millimeter or higher resolution as well as computed tomography (CT) and magnetic resonance (MR) images.
To make the Visible Human Project’s data accessible to the general public, I developed the Digital Humans CD-ROM in 1995. The intent of the product was to provide an overview of the project’s problems, results, and possible uses to the layperson. The creation of the CD-ROM, in itself, was a huge learning experience for me since I needed to master a large number of tools as well as develop a few programs from scratch.
To begin with, the data had to be segmented and rendered graphically. Segmentation required constructing a computer program to automatically label each of the billions of volume elements in both the male and female images. In order to save time and get the CD-ROM out on the market as quickly as possible, I concentrated on identifying the skin, bones, and brain of the male, and only the skin and bones of the female. After labeling the data, I designed a computer program that could project the vast amounts of information into a photo-realistic simulated view of a reconstructed body. For more technical details, you can look at a web-version of a paper presented at the Visible Human Project Conference.
TO DO Post a number of pictures and movies from the CD. Unfortunately, some of the source material used to generate the movies have been lost.
————————
Medical Visualization
Imaging technology is constantly getting better. We are able to acquire more data at higher resolutions with better delineation of human anatomy and function. But with all this information, we encounter the problem of how to use all that information productively. The answer, of course, is to make the computer process the data and give it back in a form that’s more easily understood by the medical practitioner or scientist.
As described in the 3D Digital Humans section, visualization of 3D images requires at least two steps: (1) labelling the different volume elements of the image, the process of segmentation, and (2) constructing a picture from those volume elements. (See my paper on interactive visualization for a more technical treatment.)
During my graduate studies, we sent a segmented 3D image of my head to a company with a computerized molder. The molding process, called "printing" by the company, converted the millions of image volume elements (or voxels) into a realistic life-size plastic replica of my brain, complete with an accurate depiction of the convolutions of the cortex. As a result, I may be the first person to have their brain "printed" in plastic. Luckily I'm not superstitious because my brain is sitting on a desk in the office of the vice-chairman of Neurosurgery.
————————
Artifical Intelligence
Since my early days at Stanford, I’ve been fascinated with the notion of artificial intelligence (AI). I read Ed Feigenbaum’s “Fifth Generation” on a subway in Tokyo, thinking that within a decade we’d have autonomous machines. But by the time I left Stanford, I grew disillusioned that we’d break through the “common sense” barrier. Doug Lenat has spent the last 20 years working on a solution called the Cyc project. Time will tell whether his approach is as brittle as previous mechanisms.
During my studies, I have created a software ‘robot’ for the XTank program and a self-learning Go-moku (Connect 5) program that uses artificial neural networks. Game playing provides interesting problems given the typical combinatorial explosion of possible moves, but it’s still a highly constrained playground. Also, computer solutions to game playing are very different from how the human brain handles the problem, so IBM‘s Deep Blue system is not relevant in constructing a thinking machine.
The August 2002 issue of Red Herring has an excellent article by Geoffrey James on AI hype. Fields of computer science have progressed in some areas (e.g. game playing and pattern recognition), but the end goal of AI – a truly thinking machine – remains elusive. Ray Kurzweil’s optimistic predictions (found in the highly entertaining KurzwelAI.net website) are a good contrast to the pragmatism of Terry Winograd.
Past history suggests we should make conservative predictions. In 1950, Alan Turing described a blind test to determine if a computers thinks. No computer has passed at this time. In the 1960s, computer legend says Marvin Minsky, cofounder of the MIT AI lab, gave a graduate student a simple short-term project called computer vision. We haven’t solved that problem either. In 1983, Ed Feigenbaum predicted Japan would create thinking computers and dominate the world economy. Over a decade later, the Japanese put a halt to their Fifth Generation Project without making a dent in the thinking machines problem.
Research in narrow AI fields will lead to great discoveries and practical applications. But it is questionable whether we’ll be able to replicate human thought within the next 50 years.
