A very high performing system to discriminate tissues in mammograms as benign and malignant


In this paper, we compare different stat-of-the-art texture descriptors to discriminate tissues in mammograms as either benign or malignant. The three best approaches are the following:

1) A very recent Local Ternary Pattern (LTP) variant based on a random subspace of rotation invariant bins with higher variance, where features are transformed using Neighborhood Preserving Embedding (NPE) and then used to train a support vector machine (SVM). The set of SVMs is combined by sum rule.

2) An ensemble of local phase quantization (LPQ) texture descriptors each obtained varying the parameters of LPQ. For each descriptor a SVM is trained then the SVMs are combined by sum rule.

3) A method that uses all the uniform bins extracted by LTP for training a random subspace of SVMs.

The use of these techniques is very promising when applied to the task of distinguishing benign and malignant breast tissues, with the best approach being to use all the uniform bins extracted by LTP. It obtains an area under the ROC curve (AUC) of 0.97.

Keywords: texture descriptors; local binary patterns; local ternary patterns; support vector machines; random subspace.dataset.

[full paper]