Comparative Analysis of Serial and Parallel Fusion on Texture Features for Improved Breast Cancer Diagnosis

Background: Computer Aided Diagnosis (CADx) with screening mammography could help radiologists and doctors in reliable and accurate identification of breast cancer at the preliminary stage. Objective: To propose an early diagnosis technique using fusion of texture features extracted from both Medio Lateral Oblique (MLO) and Cranio Caudal (CC) view mammograms. Methods: The proposed two-view CADx system segments the tumour region by fuzzy c-means clustering. The texture features extracted from MLO and CC view are reduced by Principal Component Analysis (PCA) or Canonical Correlation Analysis (CCA) and fused serially or parallelly followed by support vector machine classifier (SVM). Results: An improvement in accuracy of 4.4% and 7.05% was achieved with serial fusion using CCA for DDSM and INbreast datasets respectively. It is also observed that there is a significant improvement in specificity, F1 measure, kappa coefficient and Balanced Classification Rate (BCR) with serial fusion using CCA for all four texture features irrespective of the datasets. A significant improvement in BCR of 3.26% and 7.52% was achieved with LAWs feature for DDSM and INbreast respectively. Conclusion: LAWs provide almost all types of texture variations such as edges, ripples and spots. CCA transforms the feature vectors in such a way that the transformed features have maximum cross correlation and minimum auto correlation. Hence, it provides more relevant features and consequently improves the performance. Thus, our method could be used to assist doctors in enhancing the effectiveness of breast cancer diagnosis and start the treatment in earlier stage of the disease.