Research On 3D Refined Aortic Tubular Structure Semantic Segmentation Method Based On Deep Learning

Aortic vascular disease is the most threatening health killer among human cardiovascular diseases.Monitoring the health of the aorta through the morphology and structure of the aorta can promptly diagnose diseases such as aortic dissection and aortic aneurysm,which has become an important clinical computer-aided diagnosis method.Manual labeling of medical images will take a lot of time and will vary due to the subjective influence of the labeler.This paper proposes a fully automatic refined aortic segmentation algorithm based on deep learning,which effectively solves the difficult problems of medical image segmentation,especially tubular structure segmentation tasks,and has strong robustness and generalization.The first part of this paper proposes an octahedron-shaped convolution model for fusion of contextual semantic information.In order to compensate for the lack of semantic information at the aorta,especially the small diameter and bifurcation points,the octahedron-shaped convolution enhanced the features of each pixel in the feature map,and introduces the symbol distance field as a similarity measure.Octahedron-shaped convolution was divided into two sub-models: fixed octahedron-shaped convolution and adaptive octahedron-shaped convolution.The difference between the two is whether the sources of the combined features are directly adjacent in space.In addition,in this paper,the aortic shape structure is used as a constraint to introduce a radial distance loss,which is conducive to capturing a finer diameter structure.The second part of this article proposed a deep spatial convolution model based on the prior information of the continuous appearance structure of the aorta.By converting the layer-by-layer convolution in the traditional image dimension to the slice-by-slice convolution on channel,the semantic information of the image can be propagated forward across the channel in sequence.The deep spatial convolution changed the composing way of the semantic information of the image.The semantic information of the lower layer feature maps comes from the rich information in different directions in the upper layer feature maps,thereby image segmentation task is performed in a larger receptive field.The deep spatial convolution has extremely strong semantic information extraction delivery ability.This method proposed two sub-models,serial deep spatial convolution and parallel deep spatial convolution,to verify the effectiveness of the information sequential transmission mechanism in feature extraction.