Application of the Dice Similarity Coefficient (DSC) for Failure Detection of a Fully-Automated Atlas Based Knee MRI Segmentation Method

Problem Quantitative analysis of MRI images is providing new insight into and sensitivity to detect osteoarthritic progression, but is encumbered with the time, cost and variability associated with manual or semi-automated segmentation. To address this, a fully-automated knee MRI segmentation and analysis method was developed and validated. Although the method has proven to be robust, in a small percentage of cases (< 2%) underlying image quality or other anomalies may produce poor segmentation results. This study examines the feasibility of using the Dice Similarity Coefficient (DSC) as an objective, reproducible and automated method of accurately detecting segmentation failure. Methods The DSC is a measurement of spatial overlap used widely for comparing segmentation results. The DSC, which can have a value ranging from zero to one, is defined as two times the volume of the intersection between two segmentations divided by the sum of the volumes of the two segmentations. The tested hypothesis was that successful segmentations would be significantly more similar to the anatomical atlas used for generating the segmentations than failed segmentations. This is intuitive as atlas based segmentation techniques rely on predefined shapes to properly segment similar ones. Thus, we expected that DSC values calculated between the atlas and successful segmentations would be significantly higher than that of the atlas and failed segmentations.