Application of 3D U-net in automatic segmentation of middle ear surgery structures in temporal bone CT

Abstract

Objective:To study the feasibility of fully automatic segmentation of labyrinth, facial nerve and ossicles in clinical routine temporal bone CT images based on 3D U-net neural network. Method:Clinical data were divided into two groups: ①Normal group: data were randomly assigned from 30 patients for routine temporal bone CT examination; ②Abnormal group: cochlear, ossicles and facial nerve morphology variation of 1 case each. The structures of facial nerve, labyrinth and ossicles were manually initial segmented and fine segmented by 2 clinicians with Mimics 20.0. Three-dimensional convolutional neural network（3D U-Net） was selected to conduct deep learning on the same data. The dice similarity coefficient（DSC） was used as the evaluation index. Result:The 3D U-net neural network was used to automatically segment the labyrinth, ossicles and facial nerve in the routine temporal bone CT. In the normal group, the DSC of labyrinth, ossicles and facial nerve were 0.79±0.03, 0.64±0.05 and 0.49±0.09, respectively. In the abnormal group, the DSC of these structures were 0.71, 0.54 and 0.40. Conclusion:According to the anatomical characteristics of the temporal bone, the labyrinth, ossicles and the facial nerve can be totally automatic segmented by 3D U-net neural network, and the accuracy was closed to that of manual segmentation. This method is feasible, fast and accurate.