Geometric and dosimetric evaluation of atlas based auto-segmentation of cardiac structures in breast cancer patients.

BACKGROUND AND
PURPOSE: Auto-segmentation represents an efficient tool to segment organs on CT imaging. Primarily used in clinical setting, auto-segmentation plays an increasing role in research, particularly when analyzing thousands of images in the "big data" era. In this study we evaluate the accuracy of cardiac dosimetric endpoints derived from atlas based auto-segmentation compared to gold standard manual segmentation.
MATERIAL AND METHODS: Heart and cardiac substructures were manually delineated on 54 breast cancer patients. Twenty-seven patients were used to build the auto-segmentation atlas, the other 27 to validate performance. We evaluated accuracy of the auto-segmented contours with standard geometric indices and assessed dosimetric endpoints.
RESULTS: Auto-segmented contours overlapped geometrically with manual contours of the heart and chambers with Dice-similarity coefficients of 0.93 ± 0.02 (mean ± standard deviation) and 0.79 ± 0.07 respectively. Similarly, there was a strong link between dosimetric parameters derived from auto-segmented and manual contours (R2 = 0.955-1.000). On the other hand, the left anterior descending artery had little geometric overlap (Dice-similarity coefficient 0.09 ± 0.07), though acceptable representation of dosimetric parameters (R2 = 0.646-0.992).
CONCLUSIONS: The atlas based auto-segmentation approach delineates heart structures with sufficient accuracy for research purposes. Our results indicate that quality of auto-segmented contours cannot be determined by geometric values only.