3D FractalNet: Dense volumetric segmentation for cardiovascular MRI volumes

Abstract

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.