Cochlear CT image segmentation based on u-net neural network

Objective: Before cochlear implantation, accurately identifying the cochlea's morphology is necessary. This study proposes an improved network model based on U-Net, which can realize automatic segmentation of human cochlear anatomy in computed tomography (CT) images. Methods: The CT scan data of 100 patients requiring cochlear implantation diagnosed in our hospital were randomly collected. It was divided into a training set (n = 75) and a test set (n = 25). All data were manually segmented by two clinicians. At the same time, U-Net was used for deep learning of the above data. The cochlea in the test set was compared with the dice similarity coefficient (DSC) and 95% Hausdorff surface distance (HD95%) by manual and automatic segmentation. Results: The DSC and HD95% of manual cochlear image segmentation were 0.761 and 4.343, respectively. The DSC and HD95% were 0.742 and 4.217, respectively, for automatic segmentation of cochlear structure using the U-Net network structure. The difference of DSC and HD95% between the two segmentation methods was not statistically significant (P > 0.05). Conclusions: The cochlea can be thoroughly segmented automatically based on the U-Net neural network, and the precision is close to manual segmentation.