A QUANTITATIVE EVALUATION MEASURE FOR 3D BIOMEDICAL IMAGE SEGMENTATION

摘要

Modelling and quantification of biomedical data often requires segmentation as the pre-process where qualitative and quantitative evaluation measures of biomedical image segmentation play a crucial role in image analysis. The existing two-dimensional (2D) image segmentation evaluation measure and comparison techniques provide limited measurements when applied to three-dimensional (3D) image array corresponding to a stack of 2D slices and are not as practical. In this paper, we propose an approach to evaluate the quality and the accuracy of 3D image segmentation results using similarity indices based on both overlapping volume and edge based techniques. While the edge based technique was more sensitivity in detecting segmentation misclassifications, it was affected by the orthogonal data set (axial, coronal or sagittal) used for evaluation. Thus it is important for segmentation measures to consider the 3D data set as a whole, rather than a stack of independent 2D images.