Ground-Truth Segmentation of the Spinal Cord from 3T MR Images Using Evolutionary Computation

Spinal cord atrophy is one of the neuroimaging features associated with neurodegenerative diseases, inflammatory diseases and trauma. MR images segmentation can be used to assess cord atrophy, with varying degrees of manual intervention. However the accuracy of segmentation results highly depends on the operator's experience: there is a clear need for methods that simplifies and facilitates expert intervention, while providing an accurate quantification of cord atrophy. We propose and test here a ground-truth segmentation based on a simple evolutionary algorithm. EAcord integrates a set of segmentation methods with varying accuracy and manual intervention (manual, semi-automated and automated methods), as well as knowledge about the spinal cord anatomy, its relative location, immediate surrounding environment and shape at C2 vertebral level. A lighter version, EAcord-light, is also proposed, using only segmentations from semi-automated and automated methods as inputs. An experimental analysis of both algorithms showed an improved reproducibility and similar or even better accuracies compared to manual outlining. More interestingly, in some cases, EAcord-light produced a ground-truth segmentation with minimal expert intervention.