Authors:
Lucca R. Peregrino
1
;
Jordy V. Gomes
1
;
Thaís G. do Rêgo
1
;
Yuri de A. M. Barbosa
1
;
Telmo de M. e Silva Filho
2
;
Andrew D. A. Maidment
3
and
Bruno Barufaldi
3
Affiliations:
1
Center of Informatics, Federal University of Paraíba, João Pessoa, Brazil
;
2
Department of Statistics, Federal University of Paraíba, João Pessoa, Brazil
;
3
Department of Radiology, University of Pennsylvania, 3640 Hamilton Walk, Philadelphia PA, U.S.A.
Keyword(s):
Phantoms, Tomosynthesis, Deep Learning, U-Net, Segmentation.
Abstract:
Digital breast tomosynthesis (DBT) has rapidly emerged for screening mammography to improve cancer detection. Segmentation of dense tissue plays an important role in breast imaging applications to estimate cancer risk. However, the current segmentation methods do not guarantee an ideal ground-truth in clinical practice. Computer simulations provide ground-truth that enables the development of convolutional neural network (CNN) applications designed for image segmentation. This study aims to train a CNN model to segment dense tissue in DBT images simulated using anthropomorphic phantoms. The phantom images were simulated based on clinical settings of a DBT system. A U-Net, a CNN model, was trained with 2,880 images using a slice-wise approach. The U-Net performance was evaluated in terms of percent of density in the central slice and volumetric breast density in the medio-lateral slices. Our results show that the U-Net can segment dense tissue from DBT images with overall loss, accura
cy, and intersection over union of 0.27, 0.93, and 0.62 in the central slices, and 0.32, 0.92, and 0.54 in the medio-lateral slices, respectively. These preliminary results allow us to explore the use of CNN architectures to segment dense tissue in clinical images, which is a highly complex task in screening with DBT.
(More)