Evaluation

Using manual segmentation as the gold standard, we performed quantitative and qualitative evaluation of the results through automatic segmentation based on deep learning. In this study, Dice coefficient (DC) and Average symmetric surface distance (ASSD) were used as the evaluation criteria of segmentation accuracy to assess the congruence of automatic and manual segmentation [32–35]. The DC and ASSD were defined as follows:

where R is the manually outlined mask developed by the clinicians, and R₀ is the segmented mask using the deep learning method. d (r, r₀) is the Euclidian distance between the two voxels. r and r₀ are the surface points of R and R₀, respectively. N_R and N_R0 are the number of surface voxels on R and R₀, respectively. The DC is spatial overlap index [36]. ASSD serves as a metric of shape similarity, which can supplement descriptions of structural margins [35]. Higher DC values (maximum value of 1) indicate that the automatic segmentation is more similar to ground truth. For ASSD, lower values (minimum value of 0) represent greater agreement between the automatic segmentation contours and gold standard.