Image labeling and parcellation (i.e. assigning structure to a collection of voxels) are critical tasks for the assessment of volumetric and morphometric features in medical imaging data. The process of image labeling is inherently error prone as images are corrupted by noise and artifacts. Even expert interpretations are subject to subjectivity and the precision of the individual raters. Hence, all labels must be considered imperfect with some degree of inherent variability. One may seek multiple independent assessments to both reduce this variability and quantify the degree of uncertainty. Existing techniques have exploited maximum a posteriori statistics to combine data from multiple raters and simultaneously estimate rater reliabilities. Although quite successful, wide-scale application has been hampered by unstable estimation with practical datasets, for example, with label sets with small or thin objects to be labeled or with partial or limited datasets. As well, these approaches have required each rater to generate a complete dataset, which is often impossible given both human foibles and the typical turnover rate of raters in a research or clinical environment. Herein, we propose a robust approach to improve estimation performance with small anatomical structures, allow for missing data, account for repeated label sets, and utilize training/catch trial data. With this approach, numerous raters can label small, overlapping portions of a large dataset, and rater heterogeneity can be robustly controlled while simultaneously estimating a single, reliable label set and characterizing uncertainty. The proposed approach enables many individuals to collaborate in the construction of large datasets for labeling tasks (e.g., human parallel processing) and reduces the otherwise detrimental impact of rater unavailability.

Labels that identify specific anatomical and functional structures within medical images are essential to the characterization of the relationship between structure and function in many scientific and clinical studies. Automated methods that allow for high throughput have not yet been developed for all anatomical targets or validated for exceptional anatomies, and manual labeling remains the gold standard in many cases. However, manual placement of labels within a large image volume such as that obtained using magnetic resonance imaging is exceptionally challenging, resource intensive, and fraught with intra- and inter-rater variability. The use of statistical methods to combine labels produced by multiple raters has grown significantly in popularity, in part, because it is thought that by estimating and accounting for rater reliability estimates of the true labels will be more accurate. This paper demonstrates the performance of a class of these statistical label combination methodologies using real-world data contributed by minimally trained human raters. The consistency of the statistical estimates, the accuracy compared to the individual observations, and the variability of both the estimates and the individual observations with respect to the number of labels are presented. It is demonstrated that statistical fusion successfully combines label information using data from online (Internet-based) collaborations among minimally trained raters. This first successful demonstration of a statistically based approach using minimally trained raters opens numerous possibilities for very large scale efforts in collaboration. Extension and generalization of these technologies for new applications will certainly present fascinating areas for continuing research.

Labeling or parcellation of structures of interest on magnetic resonance imaging (MRI) is essential in quantifying and characterizing correlation with numerous clinically relevant conditions. The use of statistical methods using automated methods or complete data sets from several different raters have been proposed to simultaneously estimate both rater reliability and true labels. An extension to these statistical based methodologies was proposed that allowed for missing labels, repeated labels and training trials. Herein, we present and demonstrate the viability of these statistical based methodologies using real world data contributed by minimally trained human raters. The consistency of the statistical estimates, the accuracy compared to the individual observations and the variability of both the estimates and the individual observations with respect to the number of labels are discussed. It is demonstrated that the Gaussian based statistical approach using the previously presented extensions successfully performs label fusion in a variety of contexts using data from online (Internet-based) collaborations among minimally trained raters. This first successful demonstration of a statistically based approach using “wild-type” data opens numerous possibilities for very large scale efforts in collaboration. Extension and generalization of these technologies for new application spaces will certainly present fascinating areas for continuing research.