Abstract #0269

Efficient Anatomical Labeling by Statistical Recombination of Partially Label Datasets

Bennett Allan Landman¹, John Anton Bogovic², Jerry Ladd Prince^2,3

¹Biomedical Engineering, Johns Hopkins University, Baltimore, MD, USA; ²Electrical and Computer Engineering, Johns Hopkins University, Baltimore, MD, USA; ³Biomedical Engineering, Johns Hopkins University, Balitmore, MD, USA

Manual labeling of medical imaging data is critical task for the assessment of volumetric and morphometric changes; however, even expert raters are imperfect and subject to variability. Existing techniques to combine data from multiple raters require that each rater generate a complete dataset. We propose a robust extension which allows for missing data, accounts for repeated tasks, and utilizes training data. With our technique, numerous raters can label small, overlapping portions of a large dataset, and rater heterogeneity can be robustly controlled while estimating a single, reliable label set. This enables parallel processing and reduces detrimental impacts of rater unavailability.