Evaluation of multiple automatic knee cartilage segmentation algorithms

Abstract

Paperi käsittelee automaattisten segmentointialgoritmien kykyä segmentoimaan polvien nivel rustoa. Kyseisessä tutkimuksessa käsiteltiin viittä eri automaattista segmentointialgoritmia kvantitatiivisesti.

Commonly in osteoarthritis studies, large amounts of MRI data are acquired and cartilage is manually delineated from the MRI data. We investigate automatic segmentation frameworks in order to obtain quantitative data on articular cartilage morphology. We cover Mokkula, a manual segmentation framework, atlas-based automatic segmentation methodologies and a patch-based technique comparing their respective segmentation accuracies. Using Laplace’s equation to calculate cartilage thickness error (LTE), a vector thickness error method (VTE) and Dice Similarity Coefficient (DSC) to assess the accuracy of these techniques. The most accurate segmentations reached DSC of 0,87 on both the Femur and Tibia. The thickness analysis gave avarage errors of 0,32mm over the Femur and 0,36 mm over the Tibia. We feel these values are reaching high enough standards to be used in large studies.