TY - GEN
T1 - 3D FractalNet: Dense volumetric segmentation for cardiovascular MRI volumes
AU - Yu, Lequan
AU - Yang, Xin
AU - Qin, Jing
AU - Heng, Pheng Ann
PY - 2017/1/1
Y1 - 2017/1/1
N2 - Cardiac image segmentation plays a crucial role in various medical applications. However, differentiating branchy structures and slicing fuzzy boundaries from cardiovascular MRI volumes remain very challenging tasks. In this paper, we propose a novel deeply-supervised 3D fractal network for efficient automated whole heart and great vessel segmentation in MRI volumes. The proposed 3D fractal network takes advantage of fully convolutional architecture to perform efficient, precise and volume-to-volume prediction. Notably, by recursively applying a single expansion rule, we construct our network in a novel self-similar fractal scheme and thus promote it in combining hierarchical clues for accurate segmentation. More importantly, we employ deep supervision mechanism to alleviate the vanishing gradients problem and improve the training efficiency of our network on small medical image dataset. We evaluated our method on the HVSMR 2016 Challenge dataset. Extensive experimental results demonstrated the superior performance of our method, ranking top in both two phases.
AB - Cardiac image segmentation plays a crucial role in various medical applications. However, differentiating branchy structures and slicing fuzzy boundaries from cardiovascular MRI volumes remain very challenging tasks. In this paper, we propose a novel deeply-supervised 3D fractal network for efficient automated whole heart and great vessel segmentation in MRI volumes. The proposed 3D fractal network takes advantage of fully convolutional architecture to perform efficient, precise and volume-to-volume prediction. Notably, by recursively applying a single expansion rule, we construct our network in a novel self-similar fractal scheme and thus promote it in combining hierarchical clues for accurate segmentation. More importantly, we employ deep supervision mechanism to alleviate the vanishing gradients problem and improve the training efficiency of our network on small medical image dataset. We evaluated our method on the HVSMR 2016 Challenge dataset. Extensive experimental results demonstrated the superior performance of our method, ranking top in both two phases.
UR - http://www.scopus.com/inward/record.url?scp=85011266714&partnerID=8YFLogxK
U2 - 10.1007/978-3-319-52280-7_10
DO - 10.1007/978-3-319-52280-7_10
M3 - Conference article published in proceeding or book
SN - 9783319522791
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 103
EP - 110
BT - Reconstruction, Segmentation, and Analysis of Medical Images - 1st International Workshops, RAMBO 2016 and HVSMR 2016 Held in Conjunction with MICCAI 2016, Revised Selected Papers
PB - Springer Verlag
T2 - 1st International Workshops on Reconstruction and Analysis of Moving Body Organs, RAMBO 2016 and 1st International Workshops on Whole-Heart and Great Vessel Segmentation from 3D Cardiovascular MRI in Congenital Heart Disease, HVSMR 2016 Held in Conjunction with 19th International Conference on Medical Image Computing and Computer-Assisted Intervention, MICCAI 2016
Y2 - 17 October 2016 through 21 October 2016
ER -