Multi-scale volumetric convnet with nested residual connections for segmentation of anterior cranial base

Abstract

Anterior cranial base (ACB) is known as the growth-stable structure. Automatic segmentation of the ACB is a prerequisite to superimpose orthodontic inter-treatment cone-beam computed tomography (CBCT) images. The automatic ACB segmentation is still a challenging task because of the ambiguous intensity distributions around fine-grained structures and artifacts due to the limited radiation dose. We propose a fully automatic segmentation of the ACB from CBCT images by a volumetric convolutional network with nested residual connections (NRN). The multi-scale feature fusion in the NRN not only promotes the information flows, but also introduces the supervision to multiple intermediate layers to speed up the convergence. The multi-level shortcut connections augment the feature maps in the decompression pathway and the end-to-end voxel-wise label prediction. The proposed NRN has been applied to the ACB segmentation from clinically-captured CBCT images. The quantitative assessment over the practitioner-annotated ground truths demonstrates the proposed method produces improvements to the state-of-the-arts.