Two-stage multi-scale breast mass segmentation for full mammogram analysis without user intervention

Mammography is the primary imaging modality used for early detection and diagnosis of breast cancer. X-ray mammogram analysis mainly refers to localization suspicious regions interest followed by segmentation, towards further lesion classification into benign versus malignant. Among diverse types abnormalities, masses are most important clinical findings carcinomas. However, manually segmenting from native mammograms time-consuming error-prone. Therefore, an integrated computer-aided system required assist clinicians automatic precise mass delineation. In this work, we present a two-stage multi-scale pipeline that provides accurate contours high-resolution full mammograms. First, propose extended deep detector integrating fusion strategy automated localization. Second, convolutional encoder-decoder network using nested dense skip connections employed fine-delineate candidate masses. Unlike previous studies based on segmentation regions, our framework handles without any user intervention. Trained INbreast DDSM-CBIS public datasets, achieves overall average Dice 80.44% test images, outperforming state-of-the-art. Our shows promising accuracy as full-image system. Extensive experiments reveals robustness against diversity size, shape appearance masses, better interaction-free diagnosis.