An approach to computer aided diagnosis by multi-layer preceptrons

Due to lack of specificity, many computer aided breast cancer diagnosis tools entail unnecessary surgical biopsies. This study explores the discriminatory power of heterogeneous mammographic and sonographic descriptors in solving the classification task. We used various feature selection techniques to find a set of descriptors that outperforms those from other studies. Model performance was estimated by ROC analysis and metrics which show that the proposed descriptor set outperforms the others in all the metrics with most significant improvement in specificity at high levels of sensitivity.