Analysis of the performance of a multi-view fusion and active contour constraint based deep learning algorithm for ossicles segmentation on 10 μm otology CT

Abstract

Objective: To explore the performance of a deep learning algorithm that combined multi-view fusion with active contour constrained for ossicles segmentation on the 10 μm otology CT images. Methods: The 10 μm otology CT image data from 79 cases (56 cases were from volunteers and 23 cases were from specimens) were retrospectively collected in the Radiology Department of Beijing Friendship Hospital from October 2019 to December 2020. An annotation of malleus, incus, and stapes were conducted. Then the datasets were established and were divided into training set (n=55), validation set (n=8), and test set (n=16). Using the rapid localization of the region of interest combined with the precise segmentation algorithm, the malleus, incus and stapes were segmented and fused from three perspectives of coronal, sagittal and cross-sectional views. Besides, an active contour loss was designed simultaneously for the segmentation of stapes. Dice similarity coefficient (DSC) was used as the objective evaluation metric for the evaluation of the segmentation results. The inter group DSC of the proposed method was compared with that of the basic method and other methods. Results: The average DSC values of the multi-view fusion segmentation algorithm for malleus, incus and stapes reached up to 94.2%±2.7%, 94.6%±2.6% and 76.0%±5.5%, respectively. After adopting the constraint of active contour loss method, the average DSC of stapes was improved (76.4%±5.4% vs 76.0%±5.5%). The visualization results also demonstrated that the segmentation results of the stapes were more complete. Conclusions: Multi-view fusion algorithm based on 10 μm otology CT images can realize accurate segmentation of malleus and incus. Combined with the constraint of active contour loss method, the segmentation accuracy of stapes can be further improved.