Purchase this article with an account.
Li Tang, Mona K. Garvin, Young H. Kwon, Michael D. Abramoff; Segmentation of Optic Nerve Head Rim in Color Fundus Photographs by Probability Based Active Shape Model. Invest. Ophthalmol. Vis. Sci. 2012;53(14):2144.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To segment the rim of the optic nerve head (ONH), one of the major anatomical structures in fundus photographs of the retina. Accurate segmentation of the rim is essential for quantifying damage to the optic disc in glaucoma and other optic neuropathies.
Fifty-eight patients diagnosed as glaucoma suspect or primary open-angle glaucoma were included with their left eyes imaged. The ONH rim boundaries were delineated by 3 glaucoma specialists and used as a reference standard. Trained with 14 images using leave-one-out cross-validation, pixel feature classification labeled each pixel with its rim probability. An active shape model (ASM) was built from another 15 training images which had both the rim probabilities and the reference standards available. Mean optic disc shape quantified as 16 evenly spaced contour points was obtained by principal component analysis (PCA), which was first aligned with the testing image by translation, rotation and scaling according to the rim probability profile of the landmark points across rim boundaries. Residual adjustments for each individual point were transformed into a "model parameter space" so that the shape constraints were enforced through "mode of shape variations" or parameters associated with the PCs, i.e. the ONH rim boundaries only deformed into shapes consistent with those learned in training. Then the rim probability profiles of the individual point at their updated locations were compared with those in the training set and new local movements were suggested. With this iterative refinement, hard ONH rim labels were assigned to each pixel as output. The model was evaluated on 29 testing images using the percentage of pixels assigned the correct class (accuracy). The mean distance of each pixel on the estimated rim boundaries to the closest point on the ground truth was measured relative to the equivalent diameter of the optic disc (error).
The accuracy is 93.44% ± 4.22% (false positive rate: 2.22%; true positive rate: 77.87%) and the error is 7.25% ± 0.15%.
ASM based ONH rim segmentation has the potential to provide improved robustness of performance by characterizing its shape variability and incorporate it as a priori knowledge into the automated system.
This PDF is available to Subscribers Only