Probabilistic tractography of the optic radiations--an automated method and anatomical validation.

Accurately tracing the optic radiations in living humans has important implications for studying the relationship between tract structure or integrity and visual function, in health and disease. Probabilistic tractography is an established method for tracing white matter tracts in humans. Prior studies have used this method to trace the optic radiations, but operator-dependent factors, particularly variability in seed voxel placement and choice of connectivity threshold to select between tract and non-tract voxels, remain potential causes of significant variability. Methods using prior information to modify tract images risk introducing error by underestimating individual variability, particularly in subjects with abnormal anatomy. Finally, existing methods lack thorough validation against a histological standard, causing difficulty in evaluating individual methods, and quantitatively comparing methods. Here we describe a method for producing binary optic radiation images using an existing, well-validated tractography method. All stages are automated, including mask image generation, and thresholds are objectively selected by comparing tract images with existing probabilistic histological data in stereotaxic space. Data from two subject groups are presented; the first used to derive analysis parameters, and the second to test these parameters in an independent sample. Validation utilised a novel variant of receiver operating characteristic analysis, providing both justification for this method and a metric by which tractography methods might be compared generally. The resulting tracts match the histological data well; images generated in individuals matched the histological group data about as well as did images derived in individuals from that histological data set, with a low false positive rate.

Coding of the contrasts in natural images by populations of neurons in primary visual cortex (V1).

It is possible to discriminate between grating contrasts over a 300-fold contrast range, whereas V1 neurons have very limited dynamic ranges. Using populations of model neurons with contrast-response parameters taken from electrophysiological studies (cat and macaque), we investigated ways of combining responses to code contrast over the full range. One model implemented a pooling rule that retained information about individual response patterns. The second summed responses indiscriminately. We measured accuracy of contrast identification over a wide range of contrasts and found the first model to be more accurate; the mutual information between actual and estimated contrast was also greatest for this model. The accuracy peak for the population of cat neurons coincided with the peak of the distribution of contrasts in natural images, suggesting an ecological match. Macaque neurons seem better able to code contrasts that are slightly higher on average than those found in the natural environment.