Clustered Microcalcifications Candidates Detection in Mammograms

Context: Mammary microcalcifications are not-palpable lesions that are present in approximately 55 % of breast cancer. These are a frequent findings in mammograms and may be an indicator of the disease in its early stages. Method: A method was implemented in order to get mammary microcalcifications enhancement based on multi-resolution analysis with Wavelet transform. Then, candidates were segmented using threshol-ding, in this technique, the threshold was determined with statistical parameters from Wavelet distribution coefficients. Later, a couple of Support Vector Machines models was used to classify images that contains mammary microcalcifications. Results: Classification task was performed using Support Vector Machines (SVM). The following evaluation metrics was achieved: AUC of 93.6 %, accuracy of 89.4 %, sensivity of 88.4 % and specificity of 90.5 % Conclusions: In this approach the length and distribution of microcalcifications was used as features to select candidates. These features are also used as criteria in clinical evaluation to detect mammary cancer in early stages. The proposed method to image enhancement can unmask microcalcifications that are not visible at naked eye. In most mammographies the proposed algorithm classify correctly microcalcifications in different distributions.