Discrimination between tumour epithelium and stroma via perception-based features

• Tumour-stroma ratio (TSR) is an independent prognostic factor for a number of oncologic diseases.
• Manual assessment of TSR is challenging and suffers from poor intra- and interobserver agreement.
• We propose a computer-assisted approach based on perceptual features for discriminating between epithelium and stroma.
• The method can correctly differentiate epithelium from stroma with overall accuracy of ≅97%≅97%.

In this work we propose the use of image features based on visual perception for discriminating epithelium and stroma in histological images. In particular, we assess the capability of the following five visual features to correctly discriminate epithelium from stroma in digitised tissue micro-arrays of colorectal cancer: coarseness, contrast, directionality, line-likeliness and roughness. The use of features directly related to human perception makes it possible to evaluate the tissue׳s appearance on the basis of a set of meaningful parameters; moreover, the number of features used to discriminate epithelium from stroma is very small. In the experiments we used histologically-verified, well-defined images of epithelium and stroma to train three classifiers based on Support Vector Machines (SVM), Nearest Neighbour rule (1-NN) and Naïve Bayes rule (NB). We optimised SVM׳s parameters on a validation set, and estimated the accuracy of the three classifiers on a independent test set. The experiments demonstrate that the proposed features can correctly discriminate epithelium from stroma with state-of-the-art accuracy.