Breast pectoral muscle segmentation in mammograms using a modified holistically-nested edge detection network

Authors Organisations
  • Andrik Rampun(Author)
    Ulster University
    University of Sheffield
  • Karen López-Linares(Author)
    Universitat Pompeu Fabra
    Vicomtech Foundation
  • Philip J. Morrow(Author)
    Ulster University
  • Bryan W. Scotney(Author)
    Ulster University
  • Hui Wang(Author)
    Ulster University
  • Inmaculada Garcia Ocaña(Author)
    Vicomtech Foundation
  • Grégory Maclair(Author)
    Vicomtech Foundation
  • Reyer Zwiggelaar(Author)
  • Miguel A. González Ballester(Author)
    Universitat Pompeu Fabra
  • Iván Macía(Author)
    Vicomtech Foundation
Type Article
Original languageEnglish
Pages (from-to)1-17
Number of pages17
JournalMedical Image Analysis
Early online date20 Jun 2019
Publication statusPublished - 01 Oct 2019
Show download statistics
View graph of relations
Citation formats


This paper presents a method for automatic breast pectoral muscle segmentation in mediolateral oblique mammograms using a Convolutional Neural Network (CNN) inspired by the Holistically-nested Edge Detection (HED)
network. Most of the existing methods in the literature are based on handcrafted
models such as straight-line, curve-based techniques or a combination of both. Unfortunately, such models are insufficient when dealing with complex shape variations of the pectoral muscle boundary and when the boundary is unclear due to overlapping breast tissue. To compensate for these issues, we propose a neural network framework that incorporates multi-scale and multi-level learning, capable of learning complex hierarchical features to resolve spatial ambiguity in estimating the pectoral muscle boundary. For this purpose, we modified the HED network architecture to specifically find `contour-like' objects in mammograms. The proposed framework produced a probability map that can be used to estimate the initial pectoral muscle boundary. Subsequently, we process these maps by extracting morphological properties to find the actual pectoral muscle boundary. Finally, we developed two different post-processing steps to find the actual pectoral muscle boundary. Quantitative evaluation results show that the proposed method is comparable with alternative state-of-the-art methods producing on average values of 94:8 +/- 8:5% and 97:5 +/- 6:3% for the Jaccard and Dice similarity metrics, respectively, across four different databases


  • breast mammography, pectoral muscle segmentation, computer aided diagnosis, convolutional neural networks, deep learning