Optic disc segmentation for glaucoma screening system using fundus images

Figures & data

Figure 1 The flow chart for the disc segmentation algorithm.

Abbreviation: ROI, region of interest.

Figure 2 The disc segmentation procedure for TIFF images (MESSIDOR and Magrabi). (A) Localized image; (B) inpainted image; (C) segmented disc boundaries; (D) optimized disc.

Figure 3 The disc segmentation procedure for JPG images (Bin Rushed): (A) localized fundus image, (B) inpainted image, (C) first level set, (D) the localized image based on the first level set (is split from the middle), (E) second level set, (F) optimized disc contour.

Figure 4 Flowchart for the analysis of disc segmentation.

Figure 5 Examples of good disc segmentation results for both TIFF (MESSIDOR and Magrabi) and JPG (Bin Rushed) images comparing with those of six ophthalmologists.

Figure 6 Examples of bad disc segmentation results for both TIFF (MESSIDOR and Magrabi) and JPG (Bin Rushed) images comparing with those of six ophthalmologists.

Table 1 Results of disc segmentation for Bin Rushed dataset

	Area disc	Centroid disc	Both
Total number of images	195	195	195
Number of images removed due to lack of agreement among the ophthalmologists	7	15	34
Images not localized	10	10	10
Total number of images tested	178	170	151
Accuracy (number of images)	155	149	130
Accuracy (percentage)	88.2	87.6	86
Average time (seconds)	20–30

Table 2 The disc accuracy results for the six ophthalmologists and the segmentation algorithm for Bin Rushed dataset

	Ophth 1	Ophth 2	Ophth 3	Ophth 4	Ophth 5	Ophth 6	Auto
Total number of images	195	195	195	195	195	195	195
Number of images removed due to lack of agreement among the ophthalmologists	19	22	21	29	23	22	34
Images not localized	10	10	10	10	10	10	10
Total number of images tested	166	163	164	156	162	163	151
Accuracy (number of images)	134	124	115	145	137	139	130
Accuracy (%)	80.7	76	70.1	92.9	84.5	85.2	86

Table 3 Results of disc segmentation for Magrabi dataset

	Area disc	Centroid disc	Both
Total number of images	95	95	95
Number of images removed due to lack of agreement among the ophthalmologists	8	10	21
Images not localized	6	6	6
Total number of images tested	81	79	68
Accuracy (number of images)	73	71	61
Accuracy (%)	90.1	89.8	89.7
Average time (seconds)	70–120 seconds due to the large size of the images

Table 4 The disc accuracy results for the six ophthalmologists and the segmentation algorithm for Magrabi images set

	Ophth 1	Ophth 2	Ophth 3	Ophth 4	Ophth 5	Ophth 6	Auto
Total number of images	95	95	95	95	95	95	95
Number of images removed due to lack of agreement among the ophthalmologists	14	15	15	18	15	16	21
Images not localized	6	6	6	6	6	6	6
Total number of images tested	75	74	74	71	74	73	68
Accuracy (number of images)	58	56	58	65	60	69	61
Accuracy (%)	77.3	75.6	78.36	91.5	81	94.5	89.7

Table 5 Results of disc segmentation for MESSIDOR dataset

	Area disc	Centroid disc	Both
Total number of images	260	260	260
Number of images removed due to lack of agreement among the ophthalmologists	2	5	28
Images not localized	10	10	10
Total number of images tested	248	245	222
Accuracy (number of images)	208	203	179
Accuracy (%)	86.6	82	80
Average time (seconds)	20–30 seconds

Table 6 The disc accuracy results for the six ophthalmologists and the segmentation algorithm for MESSIDOR dataset

	Ophth 1	Ophth 2	Ophth 3	Ophth 4	Ophth 5	Ophth 6	Auto
Total number of images	260	260	260	260	260	260	260
Number of images removed due to lack of agreement among the ophthalmologists	14	16	23	21	21	20	28
Images not localized	10	10	10	10	10	10	10
Total number of images tested	236	234	227	229	229	230	222
Accuracy (number of images)	195	177	208	209	187	199	179
Accuracy (%)	82.6	75.6	91.6	91.2	81.6	86.5	80

Table 7 Results of disc segmentation for all three datasets combined

	Area disc	Centroid disc	Both
Total number of images	550	550	550
Number of images removed due to lack of agreement among the ophthalmologists	17	30	83
Images not localized	26	26	26
Total number of images tested	507	494	441
Accuracy (number of images)	436	423	370
Accuracy (%)	85.9	85.6	83.9

Table 8 Results of disc segmentation by the six ophthalmologists and the algorithm for all three datasets combined

	Ophth 1	Ophth 2	Ophth 3	Ophth 4	Ophth 5	Ophth 6	Auto
Total number of images	550	550	550	550	550	550	550
Number of images removed due to lack of agreement among the ophthalmologists	47	53	59	68	59	58	83
Images not localized	26	26	26	26	26	26	26
Total number of images tested	477	471	465	472	465	466	441
Accuracy (number of images)	387	357	381	419	384	407	370
Accuracy (%)	81.1	75.7	81.9	88.7	82.5	87.3	83.9

Figure 7 The disc accuracy results for all six ophthalmologists as well as the algorithm for the three datasets separately and combined.

Note: Y axis represents the accuracy (percentage).

Table 9 The number of images with disc agreement between the ophthalmologists as well as the algorithm

	Ophth 1	Ophth 2	Ophth 3	Ophth 4	Ophth 5	Ophth 6	Auto
Ophth 1	550	318	336	378	345	364	344
Ophth 2	318	550	312	356	324	336	316
Ophth 3	336	312	550	375	344	362	335
Ophth 4	378	356	375	550	383	410	379
Ophth 5	345	324	344	383	550	364	349
Ophth 6	364	336	362	410	364	550	361
Auto	344	316	335	379	349	361	550
Total	2,085	2,046	2,064	2,281	2,109	2,197	2,084

Figure 8 The disc agreement results for the six ophthalmologists as well as the algorithm for all datasets combined together.

Note: Y axis represents 550 testing images.

Figure 9 The disc agreement results for the six ophthalmologists as well as the algorithm for all datasets.

Note: Y axis represents the total number of images that were agreed.