|
KMID : 1147120170230010014
|
|
Journal of the Korean Society of Imaging Informatics in Medicine
2017 Volume.23 No. 1 p.14 ~ p.20
|
|
|
Review of Evaluation Metrics for 3D Medical Image Segmentation
|
|
Kim Jang-Woo
Kim Jong-Hyo
|
|
|
Abstract
|
|
|
Background: Although the research on the automatic medical image segmentation method is active, the research on the evaluation of the image segmentation result is insufficient. It can’t be possible to study the accurate and unbiased segmentation method without a correct evaluation of automatic segmentation result. Thus, establishment of the evaluation metrics should be prioritized. In this context, this study reviews and summarizes the 20 evaluation metrics and six classification groups suggested by Taha et al in 2015.
Materials and Methods: A total of 20 image segmentation evaluation metrics are classified into six groups as follows. The first group is the overlap-based image segmentation metrics. Sensitivity, specificity, false positive rate, false negative rate, F-measure, Dice similarity coefficient, Jaccard index and global consistency error belong to this group. The second group is a volume-based metric and has volume similarity. The third group is an information theory-based metric that has mutual information and variation of information. The fourth group is the probability based metrics composed of Interclass correlation, probability distance, Cohens kappa, and Area under ROC curve. The fifth group is a distance-based evaluation metric and reflects the spatial position of the division result. The last group is based on paircounting involving Rand index and Adjusted Rand Index.
Results: It is a reasonable to evaluate the performance of the automatic image segmentation method with minimum proper evaluation metrics. Therefore, we suggested evaluation metrics suitable for a given image segmentation problem.
Conclusion: In this study, we supplemented the lack of description of Taha’s paper and simplified it briefly. Through this review, we can perform quantitative evaluation of new image segmentation method, and use it as a basic study to analyze the segmentation problems and improve the performance of new methods to be studied later.
|
|
|
KEYWORD
|
|
|
Medical image segmentation evaluation, Evaluation metricsmetric selection
|
|
|
FullTexts / Linksout information
|
|
|
|
|
|
Listed journal information
|
|
|
|
|