An Improved Micro-Calcification Detection Algorithm Using a Novel Multifractal Texture Descriptor and CNN

Detecting individual micro-calcifications (MCs) in mammograms is a challenging problem due to heterogeneous properties and diverse composition of breast tissues. False positives (FPs) are therefore a common occurrence in the outputs of different detectors. This paper focuses on FP reduction and improvement of the final MCs detection accuracy in mammograms. The proposed method uses a combination of a MC detector which outputs a patch set containing candidate MC spots, and alpha images derived from multifractal analysis to enhance texture features of MC spots in each target patch. For further highlighting the texture features, a Weber's law based approach is used to construct a new multifractal measure and the corresponding alpha patches. In order to distinguish MC spots from the candidate set, a convolutional neural network (CNN) classifier is designed to process original mammogram patches and corresponding alpha patches together for classifying suspicious MC spots to true positive group or false positive group. Multifractal features contained in alpha images are fed into the proposed CNN model, which facilitate learning richer representations for MCs in local regions and presenting better classification performance. A digital mammogram dataset, INbreast, is used to test the proposed method. Experimental results are evaluated using free-response receiver operating characteristic (FROC) and area under the FROC curve (AUC). In our experiments, a desirable classification performance is observed after using the new alpha patch set in the designed CNN classifier, and the general MC detection results based on individual mammograms in a test set demonstrate that the proposed method reduces FP numbers and improves the MC detection accuracy effectively.