Multi-view breast thermogram analysis by fusing texture features

Breast cancer is the prevalent cancer type in women and thermography aids in premature detection by utilising ultra-sensitive infrared cameras. In this work, normal and abnormal thermograms are differentiated by extracting and fusing texture features from frontal and lateral views. Multi-view thermograms are pre-processed using anisotropic diffusion. The Region of Interest from axilla to lower breast boundary is extracted through level-set segmentation without re-initialisation. Texture features such as grey-level co-occurrence matrix, grey-level run length matrix, grey-level size zone matrix and neighbourhood grey tone difference matrix that quantitatively describe local or regional texture properties are extracted for 32-normal and 31-abnormal subjects chosen from DMR database. Using t-test, the reduced feature set is determined for frontal, right-lateral and left-lateral thermograms independently from the extracted texture features. Significant features are obtained by performing kernel principal component analysis on the reduced feature set. Feature fusion is performed on obtained significant features from frontal and lateral views to obtain a composite feature vector that is fed to least square-support vector machine employing optimised hyper-parameters to classify subjects as normal and abnormal. Experimental results indicate that fusion of texture features from frontal and lateral thermograms achieved 96% accuracy, 100% sensitivity and 92% specificity.