When Computer Vision Imitates Life
Artificial intelligence researchers at the Massachusetts Institute of Technology have designed a system for computer vision that allows a machine to recognize objects in pictures as accurately as the human brain does, even when the background is filled with visual clutter or the picture is shown sideways.
Perhaps it should be no surprise that this promising model emerged only after someone tried to reproduce what already exists in nature, the human visual system. The model’s designers endeavored to include the key neurobiological mechanisms of visual signal propagation and neural computation in the brain. (Previous systems have been guided more by the architecture of computers than of the visual system.)
But the effort to supersede the limits of earlier approaches to computer vision produced results that have been startling nonetheless, said the lead author of two recent reports about the computer-vision system.1,2 Specialized, immediate-recognition testing has shown that the software can identify objects in pictures just as often as humans do (i.e., about 80 percent of the time).
“The goal of this work is to develop a computational model of the visual cortex, one that tries to take into account all the known facts about the visual system. So the model is really a neurobiological model,” said Thomas Serre, PhD, a postdoctoral research associate in MIT’s Center for Biological and Computational Learning.
“The surprise for us was how well it worked,” Dr. Serre added. “There was no guarantee that a model that put all the biological information together would perform well. The excitement is because the system was able to deal with real-world visual situations. These aren’t bar-lines or gratings, which earlier models used.”
The researchers first “trained” their software to recognize faces and objects, and then showed it standardized picture-sets to gauge the accuracy of its identifications. The computer was as accurate as previous artificial-vision systems at detecting airplanes, motorcycles and certain facial images (95.9 to 98.2 percent accurate). However, it was 5 to 19.7 percentage points better than the earlier systems at seeing leaves, cars and a second set of faces.
After fine-tuning the software, the scientists put it up against the “gold standard,” human vision. Twenty-four volunteers were asked to determine whether there was an animal in 1,200 images flashed before them for 20 ms each. The photos ranged from close-ups, to cluttered landscapes containing a small animal. Subjects answered by pushing a yes or no button. (The computer model also “saw” the pictures for 20 ms, but the limits of processing power delayed the answers several seconds.)
The researchers kept the glimpses short in order to isolate the visual cortex’s first processing stage, which the computer model mimics, and assure comparability between the two data sets, Dr. Serre said.
“If you flash the picture fast enough, there is just enough information to let a person’s brain know what the image is, but not enough time for a neural feedback loop to become active,” he said. Often, the subjects had no conscious awareness of seeing anything. Despite this fact, the humans registered correct answers at an over- all rate of 80 percent. The artificial system scored 82 percent.
Even the studies’ senior author, longtime artificial intelligence researcher Tomaso Poggio, PhD, professor of brain and cognitive sciences at MIT, was surprised at the match between the biological and the engineered, Dr. Serre recalled. “Until this, Dr. Poggio and many others in the field would have told you that biology and computer vision are not likely to interact at any point,” Dr. Serre said. “Conversely, one could think: Evolution has been shaping the visual system for many thousands of years, so if there is one set of neurobiological parameters that is optimal in life, it should be pretty good for machines, too.”
1 Serre, T. et al. IEEE Trans Pattern Anal Mach Intell 2007;29(3):411–426.
2 Serre, T. et al. Online April 2, 2007. Proc Natl Acad Sci USA.
Clues as to Why Graves’ Disease Attacks the Eyes
An overexpressed receptor on the surface of orbital T cells appears to play a role in mobilizing and reinforcing the autoimmune processes characteristic of Graves’ disease, researchers at the University of California, Los Angeles have concluded.
The group’s findings suggest that the aberrant receptor could be used to shut the molecular door on this autoimmune process.
“If we can block the interaction between these autoantibodies and the receptor, and therefore block all the signaling consequences of the interaction, the inflammatory cascade might be ameliorated,” said Terry J. Smith, MD, the senior author on two immunology journal articles in which the group reported their results.1,2 Dr. Smith is a professor of medicine at UCLA and chief of molecular medicine at Harbor UCLA Medical Center.
In control subjects without Graves’ disease, this type of cellular receptor is present on about 15 percent of the circulating T lymphocytes in the body, the group discovered. When the receptor binds normally, to insulinlike growth factor 1 (IGF-1), more T cells are recruited to the site of an infection or injury.
In patients with Graves’ disease, the UCLA researchers discovered that there were three times the normal proportion of T cells with this receptor. Instead of attaching to IGF-1 molecules, however, these receptors bound preferentially to an autoantibody that is associated with Graves’ disease.
“For some reason in patients with Graves’ disease, the IGF-1 receptor has been recognized by immune cells as being foreign, and that elicits the antibody reaction. The immune system is saying that the IGF-1 receptor shouldn’t be there,” Dr. Smith said.
“The presence of this overexpressed IGF-1 receptor and its binding to the antibody have two consequences that we know about,” Dr. Smith said. “One is that it makes the T cells resistant to apoptosis. The other is that this type of T cell actually has a proliferative advantage, because they divide more rapidly.”
Together, these factors could account for the higher numbers of memory T cells in orbital tissue and for the chronic nature of Graves’ disease, he added.
1 Douglas, R. et al. J Immunol 2007;178(5):3281–3287.
2 Douglas, R. et al. Clin Exp Immunol 2007;148(1):64–71.
Injection Helps One Diplopic Woman See Straight
If injecting eye muscles with bupivacaine causes myotoxicity that can lead to strabismus, might this reaction be exploited as a treatment for this ocular misalignment? Three San Francisco researchers suggest the answer is yes.
Diplopia has been known for two decades to be a complication of retrobulbar anesthesia after ophthalmic surgery in adults, and the usual explanation for this has been fibrosis and contracture of the anesthetic-damaged muscle. A new case study by scientists at Smith-Kettlewell Eye Research Institute gives evidence that, instead, strabismus happens because the damaged muscle regrows larger and stronger.
Senior scientists Alan B. Scott, MD, and Joel M. Miller, PhD, and research associate Danielle E. Alexander worked with a woman who had undergone previous strabismus surgery but later developed diplopia and esotropia of 14 prism-diopters. With informed consent, they injected 4.5 ml of 0.75 percent bupivacaine into her right lateral rectus muscle.
The esotropia worsened slightly for seven days, as the muscle weakened, but then the muscle began strengthening.1 Cross-sectional MRI scans showed that the posterior portion of the muscle increased in size by 58 percent. The woman’s esotropia improved to 8 prism-diopters and her double vision ended by 16 days after the injection. At 56 days, her ocular deviation had stabilized at 4 prism-diopters. Three additional patients have been injected since, and each has had a similar positive response.
Dr. Miller credits his ophthalmologist colleague Dr. Scott—who happens to be the first researcher to use botulinum for treating strabismus—for recognizing that there must be another explanation for bupivacaine’s effect. “This appears to be the first practical way to actually increase the strength of an extraocular muscle,” Dr. Miller said. “Although strabismus surgeons talk about strengthening muscles by surgery, they can really do little more than shorten them—so that the eye’s equilibrium position is pulled toward the shortened muscle—or reposition them, so that their action on one direction is increased at the cost of reduced action in another direction.”
David L. Guyton, MD, professor of ophthalmology and head of the Krieger Children’s Eye Center at Johns Hopkins University, said he hopes cataract surgeons heed the study’s findings. “This helps explain a lot about the strabismus we’ve seen with local anesthetics after cataract and retinal detachment surgery. We’ve seen strabismus increase two- or threefold since bupivacaine became available in the early 1980s.”
Dr. Guyton noted that he and others who treat adult strabismus have been unsuccessful so far in persuading ophthalmic surgeons to switch to the sub-Tenon’s method of local anesthesia, instead of retrobulbar and peribulbar injections. “Sub-Tenon’s takes a little longer to act, but it does give quite a profound effect, without this complication,” he said.
1 Br J Ophthalmol 2007;91:146–148.
Journal Takes Studies Public
Still smarting after the Vioxx debacle, the pharmaceutical industry can submit results to a peer-reviewed journal that will document unsuccessful clinical trials, which normally aren’t published at all.
Publishing company Wiley-Blackwell says it is launching Archives of Drug Information, an open-access, online and print journal, in July in response to the clamor for better availability of information about negative and inconclusive drug trials. After Vioxx was removed from the market, journal editors learned that some studies of the drug had hinted at dangerous cardiovascular side effects, but only studies without that finding had been submitted for publication.
The NIH now requires that all clinical trials they fund be registered at www.ClinicalTrials.gov. This is optional for drug and device companies sponsoring their own studies, but forgoing registration has serious consequences.
Journals such as the New England Journal of Medicine, JAMA, Ophthalmology, American Journal of Ophthalmology and Archives of Ophthalmology have adopted a policy, put forth by the International Committee of Medical Journal Editors (ICMJE), that clinical trials will not be considered for publication unless they were fully registered with ClinicalTrials.gov “before enrollment of the first participant.” (In other words, commercial trial sponsors can’t register only the successful studies.)
A registered trial receives a unique identifier number by which one can track the study from recruitment to publication, and thus know if results are never published.