• Tidak ada hasil yang ditemukan

Adaptive Thresholding Techniques on DRIVE Database

4.4 Comparison With Existing Segmentation Methods

4.4.3 Adaptive Thresholding Techniques on DRIVE Database

The adaptive thresholding using different IDM information techniques investigated achieved average sensitivity rates of 0.7327, 0.7390, 0.7454, 0.7509, 0.7612 and 0.7644 with corresponding average accuracy rates of 0.9532, 0.9506, 0.9521, 0.9511, 0.9502 and 0.9468 respectively on DRIVE database. All the adaptive thresholding using different IDM information techniques presented significantly higher average sensitivity rates of 0.7327, 0.7390, 0.7454, 0.7509, 0.7612 and 0.7644 with significantly higher average accuracy rates of 0.9532, 0.9506, 0.9521, 0.9511, 0.9502 and 0.9468 when compared with Chaudhuri et al. [26], Jiang and Mojon [38], Zana and Klein [40], Martinez-Perez et al. [135], Niemeijer et al [35] and Yin et al. [87] with average sensitivity rates of 0.3357, 0.6399, 0.6971, 0.6389, 0.7145 and 0.6522 with their corresponding average accuracy rates of 0.8773, 0.9212, 0.9377, 0.9181, 0.9416 and 0.9267 respectively.

Five of the six adaptive thresholding techniques using different IDM information presented significantly higher average accuracy rates of 0.9532, 0.9506, 0.9521, 0.9511 and 0.9502 when compared with average accuracy rates of 0.9469, 0.9452, 0.9466, 0.9299 and 0.9461 achieved by Akram and Khan [46], Marin et al. [28], Soares et al.[23], Sz- pak and Tapamo [138] and Wang et al. [43] respectively. Three of the six adaptive thresholding techniques using different IDM information presented significantly higher average sensitivity rates of 0.7509, 0.7612 and 0.7644 when compared with the average sensitivity rate of 0.7468 achieved by Vlachos and Dermatas [41]. All the adaptive thresholding techniques using different IDM information presented significantly higher average accuracy rates of 0.9532, 0.9506, 0.9521, 0.9511, 0.9502 and 0.9468 when com- pared with the average accuracy rate of 0.9285 achieved by Vlachos and Dermatas [41].

Four of the six adaptive thresholding techniques using different IDM information pre- sented significantly higher average sensitivity rates of 0.7454, 0.7509, 0.7612 and 0.7644

when compared with Mendonca et al. [84] and Staal et al. [22] with average sensitivity rates of 0.7315 and 0.7345. Five of the six adaptive thresholding techniques using dif- ferent IDM information presented significantly higher average accuracy rates of 0.9532, 0.9506, 0.9521, 0.9511 and 0.9502 when compared with Mendonca et al. [84] and Staal et al. [22] with average accuracy rates of 0.9463 and 0.9442. Ricci and Perfetti [20]

presented no average sensitivity but a higher average accuracy rate of 0.9595 when compared with the average accuracy rates of all the adaptive thresholding techniques based on IDM information ranging from 0.9468 to 0.9532. Although Xiao et al. [83]

presented a higher average accuracy rate of 0.9529 when compared with five of the six investigated adaptive thresholding techniques based on IDM information, there were only significant differences when compared with three average accuracy rates. Two of the six adaptive thresholding using different IDM information techniques presented significantly higher average sensitivity rates of 0.7612 and 0.7644 when compared with the average sensitivity rate of 0.7513 achieved by Xiao et al. [83]. The human observer presented a significantly higher average sensitivity rate of 0.7761 when compared with all the average sensitivity rates of 0.7327, 0.7390, 0.7454, 0.7509, 0.7612 and 0.7644 achieved by the adaptive thresholding techniques based on IDM information. The higher detection rate of both large and thin vessels by the human observer [131] re- sulted in the higher average sensitivity rate achieved by the human observer [131]. Five of the six adaptive thresholding techniques based on IDM information present higher average accuracy rates of 0.9532, 0.9506, 0.9521, 0.9511 and 0.9502 when compared with the average accuracy of 0.9473 achieved by the human observer with significant differences.

The adaptive thresholding using different ASM range information techniques in- vestigated achieved average sensitivity rates of 0.7313, 0.7375, 0.7397, 0.7560, 0.7632 and 0.7650 with corresponding average accuracy rates of 0.9511, 0.9503, 0.9488, 0.9477, 0.9461 and 0.9449 respectively on DRIVE database. All the adaptive thresholding us- ing different ASM range information techniques presented significantly higher average sensitivity rates of 0.7313, 0.7375, 0.7397, 0.7560, 0.7632 and 0.7650 with significantly higher average accuracy rates of 0.9511, 0.9503, 0.9488, 0.9477, 0.9461 and 0.9449 when compared with average sensitivity rates of 0.3357, 0.6399, 0.6971, 0.6389, 0.7145 and 0.6522 with their corresponding average accuracy rates of 0.8773, 0.9212, 0.9377, 0.9181, 0.9416 and 0.9267 achieved by Chaudhuri et al. [26], Jiang and Mojon [38], Zana and Klein [40], Martinez-Perez et al. [135], Niemeijer et al [35] and Yin et al. [87]

respectively. Three of the six adaptive thresholding techniques using different ASM in- formation presented significantly higher average sensitivity rates of 0.7560, 0.7632 and

0.7650 when compared with the average sensitivity rate of 0.7468 achieved by Vlachos and Dermatas [41]. All the adaptive thresholding techniques using different ASM in- formation presented significantly higher average accuracy rates 0.9511, 0.9503, 0.9488, 0.9477, 0.9461 and 0.9449 when compared with the average accuracy rate of 0.9285 achieved by Vlachos and Dermatas [41].

Szpak and Tapamo [138] presented no average sensitivity but a significantly lower average accuracy rate of 0.9299 when compared with all the average accuracy rates of 0.9511, 0.9503, 0.9488, 0.9477, 0.9461 and 0.9449 achieved by adaptive thresholding using different ASM range information. Four of the six adaptive thresholding using different ASM range information techniques presented higher average accuracy rates of 0.9511, 0.9503, 0.9488 and 0.9477 when compared with an average accuracy rate of 0.9452 achieved by Marin et al. [28] with significant differences. Three of the six adap- tive thresholding using different ASM range information techniques presented higher average accuracy rates of 0.9511, 0.9503 and 0.9488 when compared with the average accuracy rate of 0.9469 achieved by Akram and Khan [46] with significant differences.

Mendonca et al. [84] presented a lower average sensitivity rate of 0.7315 when compared with five of the six average sensitivity rates of 0.7375, 0.7397, 0.7560, 0.7632 and 0.7650 achieved by proposed adaptive thresholding techniques based on ASM information with significant differences. Four of the six proposed adaptive thresholding techniques based on ASM information presented higher average accuracy rates of 0.9511, 0.9503, 0.9488 and 0.9477 when compared with average accuracy rates of 0.9466 and 0.9463 presented in [23] and [84] with significant differences.

Staal et al. [22] presented a lower average sensitivity rate of 0.7345 when compared with five average sensitivity rates 0.7375, 0.7397, 0.7560, 0.7632 and 0.7650 achieved by proposed adaptive thresholding techniques based on ASM information with signif- icant differences. A lower average accuracy rate of 0.9442 was also presented in [22]

when compared with five average accuracy rates of 0.9511, 0.9503, 0.9488, 0.9477, and 0.9461 achieved by the adaptive thresholding techniques based on ASM information with significant differences. Wang et al. [43] presented no average sensitivity but a lower average accuracy rate of 0.9461 when compared with four average accuracy rates 0.9511, 0.9503, 0.9488 and 0.9477 achieved by adaptive thresholding techniques based on ASM information with significant differences. Ricci and Perfetti [20] presented no average sensitivity but a higher average accuracy rate of 0.9595 when compared with all the average accuracy rates of the adaptive thresholding techniques based on ASM information. Xiao et al. [83] presented a higher average accuracy rate of 0.9529 when compared with all the average accuracy rates of the adaptive thresholding techniques

based on ASM information but a lower average sensitivity rate of 0.7513 when compared with three average sensitivity rates 0.7560, 0.7632 and 0.7650 of the six proposed adap- tive thresholding techniques based on ASM information with significant differences.

The human observer presented a significantly higher average sensitivity rate of 0.7761 when compared with all the average sensitivity rates of 0.7313, 0.7375, 0.7397, 0.7560, 0.7632 and 0.7650 achieved by the adaptive thresholding techniques based on ASM information. Three of the proposed adaptive thresholding techniques based on ASM information presented higher average accuracy rates of 0.9511, 0.9503 and 0.9488 when compared with the average accuracy rate of 0.9473 achieved by the human observer with significant differences.