DPA-DenseBiasNet: Semi-supervised 3D Fine Renal Artery Segmentation with Dense Biased Network and Deep Priori Anatomy

Abstract

3D fine renal artery segmentation on abdominal CTA image targets on the segmentation of the complete renal artery tree which will help clinicians locate the interlobar artery’s corresponding blood feeding region easily. However, it is still a challenging task that no one has reported success due to the large intra-scale changes, large inter-anatomy variation, thin structures, small volume ratio and limitation of labeled data. Hence, in this paper, we propose a novel semi-supervised learning framework named DPA-DenseBiasNet for 3D fine renal artery segmentation. The dense biased connection method is presented for multi-receptive field feature maps merging and implicit deep supervision [5] which enable the network to adapt to large intra-scale changes and improve its training process. The dense biased network (DenseBiasNet) is designed based on this method. We develop deep priori anatomy (DPA) for semi-supervised learning of thin structures. Differ from other semi-supervised methods, it embeds priori anatomical features to segmentation network which avoids inaccurate results sensitive to thin structures as optimizing targets, so that the network achieves generalization of different anatomies with the help of unlabeled data. Only 26 labeled and 118 unlabeled images were used to train our framework and it achieves satisfactory results on the testing dataset. The mean centerline voxel distance is 1.976 which reduced by 3.094 compared to 3D U-Net. The results illustrate that our framework has great prospects in the diagnosis and treatment of kidney disease.