Segmentation of Endothelial Cell Boundaries of Rabbit Aortic Images Using a Machine Learning Approach

Figure 4

Mapping of neighbourhoods of points into the components of a 9-element feature vector, [𝑓₁,𝑓₂,…,𝑓₉]. Four candidate points are considered, three corresponding to true boundary points, and one corresponding to an isolated noise pixel (B) that has an intensity similar to those of the boundary points. Note that if we plot where the points in a three-dimensional subspace of 9-dimensional space lie (red triangles), we find that pixels A and B fall in the same place in the three dimensional subspace formed by components 𝑓₄,𝑓₅, and 𝑓₆; this means that this subspace cannot be sufficient to classify pixels correctly. Creating a feature vector containing all 8 neighbouring pixels' intensities, and that of the candidate pixel, provides a much better chance of successfully teaching an SVM to recognise boundary pixels from non-boundary pixels, but even this is not sufficient for low signal-to-noise conditions (see text for details).