Optic Disc Segmentation

Techniques

• Pixel Classification
• Probability Map Computation
• Feature Selection
• Cost Function Optimization
• Convex Shape Segmentation
• Simultaneous Cup/Disc Segmentation
• Combining Region-based and Edge-based Information

Pixel Classification and Probability Maps

We utilize k-NN nearest neighbor classification on feature vectors to compute probability maps representing membership in each region within the image. Pixel classification has the advantage that it can incorporate large amounts of information about a particular pixel into a classification.

Examples of different probability maps are shown below.

Feature Selection

Feature selection techniques are used to reduce the feature set from 250+ features down to only the 6 most signification features used for detecting the optic disc. We use an automated tool to generate many different low-level features on a single image such as Gaussians at different scales, edge detectors, etc. The features are computed on the HSV, RGB and Color Opponency (RG, BY, Brightness) planes.

Cost Function Optimization

Optimization of the cost function involves the selection of the most significant features from a set. Relevant features are determined by feature selection which optimizes the cost functions based on the difference between the computer segmentation and the reference standard. Our optimization pipeline is unique because the error metric used in feature computation involves the final output of the graph search, not simply the differences in classifications.

Images below show, from left to right, the original image, the computer segmentation using our algorithm and the reference standard.

Publications