Purchase this article with an account.
J. Lowell, A. Hunter, D. Steel, M. Habib; Lesion Segmentation Evaluation . Invest. Ophthalmol. Vis. Sci. 2006;47(13):985.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To evaluate the performance of four segmentation approaches: fuzzy C–Means clustering, recursive region growing, adaptive recursive region growing, and a colour discriminant function.
The relatively small size of diabetic lesions in standard 760 by 570 fundus images can affect the accuracy of boundary demarcation. Absolute delimiting is therefore difficult to obtain. For this reason, the accuracy and precision of the four segmentation algorithms were tested using 100 gold standard lesion boundaries obtained from 27 high–resolution fundus images (3300 by 2600 pixels) with a 45–degree field of view. Lesions of varying sizes and contrast were selected to form a representative lesion collection, extracted from fundus images with different degrees of retinopathy.An ophthalmologist manually delimited lesion boundaries by depicting each lesion edge pixel. Boundaries were down–sampled by a factor of 4 to produce gold standard lesion boundaries with sub–pixel accuracy. It is against this benchmark that the accuracy and precision of the algorithm is measured.
The benchmark comparison with the aforementioned techniques was achieved by measuring the number of common pixels shared with the gold standard lesions and the algorithm's segmented area. The four algorithms under or over segment diabetic lesions with varying degrees. With the Recursive region growing algorithm, the limited intensity threshold between seed and candidate region pixels is so small (10 pixels) that only partial segmentation is possible before the intensity difference reaches its threshold. The colour and intensity difference between lesion and background retina is commonly marginal with little contrast between the two. Consequently the fuzzy, adaptive and discriminant algorithms that do not utilise the limited edge strength, and tend to overestimate the lesion boundary.
Four lesion segmentation algorithms have been compared. The large number of incorrect pixel classifications shows that there is still scope for further lesion segmentation improvement.
This PDF is available to Subscribers Only