Is it possible to diagnose autism by brain imaging?

Abstract

Harvard University Schools of Medicine & Public Health, MA, USA and Neurostatistics Laboratory, McLean Hospital, MA, USA nlange@hms.harvard.edu The root biological causes of autism spectrum disorder – autism – remain a complex mystery despite 70 years of effort to solve the puzzle. A total of 6 years after Leo Kanner first identified ‘early infantile autism’ in 1943 through astute observation of 11 children [1], he retracted his initial biological theory. Since then, no alternative theory has yet been validated by genetics or molecular biology to take its place [2]. Today’s autism research is awash with an abundance of statistically significant brain imaging results that provide little differentiation between individuals with the disorder and typically developing individuals [3], and this body of findings cannot yet be employed as useful diagnostic criteria in clinical practice. As we know, and as the Oxford English Dictionary reminds us, any image is merely “an artificial imitation or representation of the external form”, a “copy,” “likeness,” “similitude,” “semblance,” “appearance” and “shadow.” Some researchers today have exclaimed, “science is imaging”. Regardless of how ‘real’ the everimproving spatiotemporal image resolution, brilliance of color and attraction of animation appear to be, this statement broadcasts a clear misconception. Much of science is measurement, and we continue to have difficulty measuring the same brain structure twice within a short time interval and getting the same number, even after accounting for the ‘regression to the mean’ effect. The reliability of autism imaging is a major present-day limitation of the technology. MR brain images are perhaps the most beautiful of all image types (so beautiful because so mutable?), and these beauties are sometimes subjected to ‘circular analysis’ in attempts to ground their impressions in solid statistics [4]. MR images, these quantized imitations of reality, are indeed measurements amenable to statistical regression and classification analysis, but they cannot yet