2D Fast Vessel Visualization Using a Vessel Wall Mask Guiding Fine Vessel Detection
Figure 4
Comparative ROC curves between three full scans and three VWP-based scans: (a) case “2” vasculature with drusen, (b) for the more homogenous background vasculature in “1”, (c) extended drusen as in “3”. The homogenous vasculature in “b” gives a more straight relationship between sensitivity and specificity as a result of a uniform spatial distribution of FP’s and TP’s that does not allow relatively large variations as our threshold and VP pairing area radius increase. As for the guided scans the most notable observation is that SE is increasing less rapidly in the full scan case as observed in full scan graphs (d), (e) and (f) than in guided scan graphs (a), (b), and (c) as a result of faster TP detection offered a prior from VWP map. The Graphs (g), (h), and (i) show the results using double sided dual mode GS-MF on same vasculatures as those for which (a), (b), and (c) stand for. In Graphs (j), (k), and (l), ROC graphs are shown for the RG detection technique, the CL-based technique and three seed point technique for every branch of the vasculature greater in length than 10 pixels. All non VWM-guided techniques (g)(l) have a steeper ROC for small specificity (SP) values. This shows when the number of FN is lowered the improvement/decrease of the FP is proportionally larger than with VWM guided detection. Our method directly addresses the FP decrease problem. Moreover, when we use more elaborated filters [spatially modulated dual mode double sided] the performance variation after the critical point where SE reaches a local peak (at ) is significantly lower than with RG, CL-based on 3 seed point.
