Quantitative analysis of micro-calcifications for breast cancer via wavelet transform and neural network

A high-sensitivity computer-aided diagnosis algorithm which can detect and quantify micro-calcifications for early-stage breast cancer is proposed in this research. The algorithm can be divided into two phases: image reconstruction and recognition on micro-calcification regions. For Phase I, the suspicious micro-calcification regions are separated from the normal tissues by wavelet layers and Renyi's information theory. The Morphology-Dilation and Majority Voting Rule are employed to reconstruct the scattered regions of suspicious micro-calcification. For Phase II, total 49 descriptors which mainly includes shape inertia, compactness, eccentricity and grey-level co-occurrence matrix are introduced to define the characteristics of the suspicious micro-calcification clusters. In order to reduce the computation load, principal component analysis is used to transform these descriptors to a compact but efficient expression by linear combination method. The performance of proposed diagnosis algorithm is verified by intensive experiments upon realistic clinic patients. The efficacy of back-propagation neural network classifier exhibits its superiority in terms of high true positive rate(TP rate) and low false positive(FP rate) rate, in comparison to Bayes classifier.