Tumor Classification in Histological Images of Prostate Using Color Texture

We present a wavelet-based color texture approach to tumor classification in the histological images of prostate. We extend our previous work on intensity images to incorporate color information and rotational invariance. Our results on a set of 367 images stained using hematoxylin and eosin indicate that incorporating color and rotational invariance into the features significantly reduces the classification error. We obtained a 5- fold cross-validation error of 8.7% for intensity images and no rotational invariance. Incorporation of color and rotational invariance lowered the error to 4.4%, using the CIELAB space. Both results were obtained using support vector machine classifiers along with the linear kernel. The improvement achieved in classification accuracy corresponds to a significance level of 0.0093.