Comparison between imaging techniques

Since the algorithm presented in this paper is able to segment both red-free and fluorescein images, a second validation was undertaken to assess the confidence of automatic measurements in the most common clinical red-free images against the automatic measurements from the better quality fluorescein images which are normally considered to be more accurate [7]. 16 randomly taken individual bifurcations were automatically measured from red-free and fluorescein paired images. Figure 13 shows the same bifurcations of Figure 12 after the fluorescein dye injection.

Figure 13: Some examples of individual bifurcations in gray scale fluorescein images. (a - c) are the original subimages and (d - f) are the segmented and measured segments. Subimages are taken from gray scale images of size $2400 \times 2800$ pixels.

For this comparison, $X_a$ corresponds to red-free measurements and $X_b$ corresponds to fluorescein. Table II summarises the results.

Table II:

It can be seen that automatic measurements of both features in fluorescein and red-free images are not significantly different. This suggests that automatic measurements taken from the more clinically common red-free images are comparable to those from fluorescein angiographs. This result is contrary to previous manual measurements  [26] where significant differences in diameters between red-free and fluorescein were found, this result is also shown in Table II. In this case the manual normalised differences of diameters from fluorescein images are larger than in red-free.

Figure 14 shows the normalised differences $\Delta d$ plotted against the average for both automatic (a) and manual (b). The means are shown as solid lines, $\pm$ two standard deviations as dashed lines.

Figure 14: Fluorescein against red-free comparisons. Normalised differences $\Delta d$ plotted against the average for both automatic (a) and manual (b). The means are shown as solid lines, $\pm$ two standard deviations as dashed lines.

It can be seen from Figure 14(b) that there is a certain tendency for fluorescein manual measurements of diameters to be generally smaller than red-free for small vessel widths, whereas they are larger for larger vessel widths. This bias with vessel size is not observed in the automatic measurements (Figure 14(a)) and it may explain the differences of significance for the p values for manual measurements.