Prediction and location of malignant nodules and microcalcifications in mammography via Deep Learning

ObjectivesTo propose a tool to detect and locate malignant nodules and microcalcifications in mammography and judge its potential as a screening tool. MethodsIn this institutional review board approved retrospective study we presented a new tool based on deep learning techniques to predict and locate lesions in mammograms, called quantusMM. 3,114 mammograms from 976 patients were collected from Onkologikoa (Instituto Oncologico de Kutxa) databases for this purpose: 1,248 images with malignant nodules, 736 with malignant microcalcifications and 1,131 without any suspicious findings. The proposed methods split the images in patches to be able to locate the lesions in the image. Then, these methods select the patches most likely to have a lesion based on the brightness values of the pixels. 80% of the selected patches (with the corresponding outcome) were used to train deep learning algorithms and the remaining 20% were used to test the performance to classify into malignant parts or control parts. ResultsThe proposed methods obtain an area under the ROC curve (AUC) of 95.5% to predict malignant nodules using the patches, and 90.4% to predict malignant nodules into the whole images. To predict malignant microcalcifications the method obtains an AUC of 99.0% into patches and 90.0% into the whole images. ConclusionsThe proposed tool shows potential to predict and locate malignant nodules and microcalcification lesions in mammography. This new approach could help in the first screening of patients and also could greatly benefit radiologists to support decision making.