These features are based on models of light transmission and reflectance. We verified our models by
fabricating phantom (realistic imitations) blood vessels with known plaque types and checking the estimates against measured values in these cases.
The Plaque-Type Classifier
After extracting features from pixels in our IVOCT
images, we then train a support vector machine
(SVM) (Cristianini and Shawe-Taylor 2000) for classification of the individual pixels. The SVM is a state-of-the-art classification method. It is theoretically
well founded and robust to noise in the data, which
is a desirable property.
A second desirable property of the SVM is its ability to construct nonlinear classifiers through the use
of kernel functions. A kernel function implicitly
maps the input data to a possibly high dimensional
space, where it learns a linear classifier. Since this
mapping is done implicitly (that is, we never actually construct the high-dimensional feature vector), the
procedure is computationally efficient. In our work,
we use a radial basis function (RBF) kernel, which is a
commonly used kernel.
The SVM is a binary classifier. Given that we are
interested in classifying three different plaque types,
we use a one-versus-rest (OVR) approach for multi-class classification. This produces three binary classifiers, one treating each class as positive and the oth-
Figure 4. Appearance of Plaque Types in Clinical Images.
A is fibrous, B is lipid, and C is calcium. D shows the appearance of a normal blood vessel wall, which has layered structure.