Purchase this article with an account.
Mohammad Saleh Miri, Michael Abramoff, Kyungmoo Lee, Meindert Niemeijer, Andreas Wahle, Young Kwon, Mona Garvin; A Multimodal Machine-Learning-Based Approach for Segmenting the Optic Disc and Cup in Fundus and SD-OCT Images. Invest. Ophthalmol. Vis. Sci. 2013;54(15):5498.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Many existing approaches exist for segmenting the optic disc and cup from a single imaging modality. The purpose of this work is to determine the extent that simultaneously using complementary information from fundus photographs and SD-OCT volumes can enable a more accurate optic disc and cup segmentation.
Seventy SD-OCT scans (200×200×1024 voxels; voxel size 30×30×2 μm) centered at optic nerve head from 35 glaucoma patients were acquired using Cirrus HD-OCT (Carl Zeiss Meditec, Inc., Dublin, CA). Stereo color fundus photograph pairs (4096×4096 pixels; Nidek, Newark, NJ) were acquired on the same day from the same patients. After segmenting the intraretinal surfaces on each SD-OCT volume using a graph-theoretic approach, 10 features were obtained at each projected location based on 2D layer-based projection images and distances between segmented surfaces. The fundus photographs were registered to each corresponding SD-OCT volume and, using a Gaussian filter bank and distance-based features, 14 features were extracted at each pixel location. A multimodal feature set (24 total features) was created by combining the features from both modalities. A random-forest classifier was applied to both the unimodal (OCT only) feature set and multimodal feature set to classify the pixels into background, rim area, and cup classes. A reference standard was obtained by taking the majority vote of three expert-defined segmentation results from the fundus photographs.
As also shown in Table 1, the overall accuracy of multimodal approach (97.43±0.58) was significantly better than the unimodal approach (94.97±0.76) using a paired t-test (p < 0.05). A sample pixel classification result from the unimodal and multimodal approach is shown in Fig. 1.
Using complementary information from both fundus photographs and SD-OCT volumes can enable better segmentations of the optic disc and cup over that obtained using unimodal information.
This PDF is available to Subscribers Only