PSO Model Selection for False Positive Reduction in Mammography

Screening mammography is currently the most efficient procedure for early detection of breast cancer. However, interpreting mammograms by radiologists is a difficult task, which produces many false positives and so many unnecessary breast biopsies.
Several false positive reduction (FPR) algorithms are being developed to improve early detection of breast cancer. This project tackled two problems related to FPR in mammography. Firstly, texture feature extraction by evaluating different approaches including multiscale first and second order statistics of wavelet coefficients, and Haralick texture features. The second problem was optimizing the classification performance. For this, a model selection approach based on particle swarm optimization (PSO) was adopted to select the most discriminative textural features and to optimize hyper-parameters and parameters selection for a nonlinear SVM classifier.
The proposed multiscale texture features and PSO model selection were evaluated using 315 mammogram regions obtained from the mini-MIAS public dataset. Although results obtained so far demonstrated the efficacy of wavelet based texture features and PSO model selection, further validation using larger mammogram datasets is needed.