Abstract

Background: Cancer that develops in any of the parts of the breast is known as breast cancer. To reduce the breast cancer fatality, Computer Aided Diagnosis can assist physicians for accurate diagnosis.

Objective: To improve the performance of breast cancer diagnosis using fusion of local binary and ternary patterns from ultrasound B mode and elastography images.

Methods: In this work, the fusion of texture features of more sensitive B mode and more specific elastography images is suggested for breast cancer diagnosis. Both the images are speckle removed and segmented for detecting the tumorous region. Local Binary Pattern (LBP) and Local Ternary Pattern (LTP) features are extracted from both images and reduced by Principal Component Analysis or Canonical Correlation Analysis (CCA). Serial or parallel fusion is applied on the reduced features followed by Support Vector Machine classifier.

Results: Use of LTP provides better accuracy (94.7% for B mode and 93.8% for Elastography) than existing systems. The accuracy is further increased to 96.5% for LBP and 98.2% for LTP with serial fusion using CCA. This shows an improvement of 1.9% and 3.7% with respect to the B mode system with LTP. As specificity reaches 100%, no false positives have been detected.

Conclusion: LTP is more immune to noise and CCA transforms the LTPs of the two images so that the transformations have maximum cross correlation and minimum auto correlation. This provides more relevant features and improves the performance. Hence, the proposed system could assist doctors in diagnosing the breast tumors in better way.

Keywords: Breast cancer, feature fusion, PCA, CCA, ultrasound B mode, EI image.

Graphical Abstract