Breast tissue segmentation from x-ray radiographs

TY - JOUR

T1 - Breast tissue segmentation from x-ray radiographs

AU - Chen, Chen

AU - Nielsen, Mads

AU - Karssemeijer, Nico

AU - Brandt, Sami Sebastian

PY - 2014/5/21

Y1 - 2014/5/21

N2 - In this paper, we propose a robust and accurate method that segments mammograms to three distinct regions: breast tissue, pectoral muscle and background. Our approach is built around a neural, two-layer committee machine. On the first layer, individual experts, each formed by a feature vector and a classifier, vote the local class label of the mammogram. The votes are given as an input, together with a prior map, to the second layer of the committee machine, which combines the inputs by a gating network. As the first layer features, we use effective, well-known local features based on image intensity, intensity histograms, local binary patterns, and histograms of oriented gradient. As with the first-layer classifiers and the gating network, we use support vector machines. Our experiments on a database of 495 mammograms, divided into independent training, validations and test subsets, show that our method is able to segment the breast tissue without failure, and it challenges the manual expert segmentation in the level of accuracy.

AB - In this paper, we propose a robust and accurate method that segments mammograms to three distinct regions: breast tissue, pectoral muscle and background. Our approach is built around a neural, two-layer committee machine. On the first layer, individual experts, each formed by a feature vector and a classifier, vote the local class label of the mammogram. The votes are given as an input, together with a prior map, to the second layer of the committee machine, which combines the inputs by a gating network. As the first layer features, we use effective, well-known local features based on image intensity, intensity histograms, local binary patterns, and histograms of oriented gradient. As with the first-layer classifiers and the gating network, we use support vector machines. Our experiments on a database of 495 mammograms, divided into independent training, validations and test subsets, show that our method is able to segment the breast tissue without failure, and it challenges the manual expert segmentation in the level of accuracy.

U2 - 10.1088/0031-9155/59/10/2445

DO - 10.1088/0031-9155/59/10/2445

M3 - Journal article

C2 - 24778348

SN - 0031-9155

VL - 59

SP - 2445

EP - 2456

JO - Physics in Medicine and Biology

JF - Physics in Medicine and Biology

IS - 10

ER -