To quantify the shape of the segmented MAs, we compute three important MA morphological indexes (“largest caliber” [
LC], “narrowest caliber” [
NC], and the “body-to-neck ratio” [
BNR]) as defined in.
15 Specifically, we first compute the MA skeleton (or medial-axis) for every single MA using the “Scikit-image” package,
47 and then apply the Euclidean distance transformation to compute the medial radius distances
D = {
di|
i = 1, 2
, ..., N } (
N is the number of points on the MA skeleton) from all points of the MA skeleton to the background pixels (i.e., pixels of each MA contour). Consequently, the
LC value for each single MA corresponds to twice of the largest distance value in the sorted distance list
Dsorted. Due to the varied vessel lengths of different MAs, we calculate the NC value for each single MA by selecting the 10 smallest medial radius distances from
Dsorted, and double the average medial radius distances as the final
NC value. Based on the
LC and
NC values, the BNR value for each MA can be computed by using
LC/NC.
Figure 6 gives some examples of the LC and NC quantification results for the segmented MAs predicted by AOSLO-net and nnUNet trained with three different MA masks (normal, short and thick mask sets). From
Figures 6a and
6c, we can find that the AOSLO-net trained with thick MA ground truth masks attains the best MA segmentation performance; the NC quantification results (red curve) as shown in the third row of
Figure 6c are very close to the reference NC values (black dashed line) obtained from the thick masks. More details about the corresponding examples of the enhanced MA perfusion maps and the MA segmentation results using AOSLO-net and nnUNet are, respectively, shown in columns 1 to 3 of
Figure 6b. AOSLO-net can effectively detect the important small vessels for the heterogeneous MAs (see the second column in
Figure 6b), which play a very important role in different downstream tasks, (e.g., MA morphological parameter quantification [NC, BNR, convexity], MA severity stratification, and in hemodynamics simulations). The three blue dots with very high NC values in different rows of
Figure 6c are the same MAs for which nnUNet fails to detect the small vessels.