AUTHOR=Forbes Jessica L. , Kim Regina E. Y. , Paulsen Jane S. , Johnson Hans J. TITLE=An Open-Source Label Atlas Correction Tool and Preliminary Results on Huntingtons Disease Whole-Brain MRI Atlases JOURNAL=Frontiers in Neuroinformatics VOLUME=10 YEAR=2016 URL=https://www.frontiersin.org/journals/neuroinformatics/articles/10.3389/fninf.2016.00029 DOI=10.3389/fninf.2016.00029 ISSN=1662-5196 ABSTRACT=

The creation of high-quality medical imaging reference atlas datasets with consistent dense anatomical region labels is a challenging task. Reference atlases have many uses in medical image applications and are essential components of atlas-based segmentation tools commonly used for producing personalized anatomical measurements for individual subjects. The process of manual identification of anatomical regions by experts is regarded as a so-called gold standard; however, it is usually impractical because of the labor-intensive costs. Further, as the number of regions of interest increases, these manually created atlases often contain many small inconsistently labeled or disconnected regions that need to be identified and corrected. This project proposes an efficient process to drastically reduce the time necessary for manual revision in order to improve atlas label quality. We introduce the LabelAtlasEditor tool, a SimpleITK-based open-source label atlas correction tool distributed within the image visualization software 3D Slicer. LabelAtlasEditor incorporates several 3D Slicer widgets into one consistent interface and provides label-specific correction tools, allowing for rapid identification, navigation, and modification of the small, disconnected erroneous labels within an atlas. The technical details for the implementation and performance of LabelAtlasEditor are demonstrated using an application of improving a set of 20 Huntingtons Disease-specific multi-modal brain atlases. Additionally, we present the advantages and limitations of automatic atlas correction. After the correction of atlas inconsistencies and small, disconnected regions, the number of unidentified voxels for each dataset was reduced on average by 68.48%.