Weakly-Supervised End-to-End Framework for Pixel-Wise Description of Micro-Calcifications in Full-Resolution Mammograms

Breast cancer is the most prevalent cancer among women and catching the early signs of the disease is crucial for increasing patient survival rates. Mammography (MG) permits the visualization of the smallest abnormalities, such as micro-calcifications (MCs). MCs may be an early sign of breast cancer, they appear as white blobs on MG and may have various shapes and different distributions. The risk of malignancy is determined according to these parameters. We propose an end-to-end pipeline structured around a deep convolutional neural network backbone inspired by UNet3+ architecture. It aims to classify and segment benign and malignant MCs at several scales. This backbone is trained in a weakly-supervised end-to-end manner by using negative mammograms augmented with artificial calcifications. Images are represented by a stack of patches, preserving the resolution and keeping MC pixel information. A post-processing stage based on a set of morphological operations is applied to clean the segmentation masks. This stage is designed with clinical knowledge about MCs and is intended to clear false activations without affecting the detection sensitivity. We tested our pipeline on two public datasets INBreast and Breast Micro-calcifications Dataset and achieved a true positive rate of 0.95 and 0.92 at the level of lesions, and 0.95 and 0.94 image-wise.