Automatic segmentation of temporal bone structures from clinical conventional CT using a CNN approach

Abstract

Background Automatic segmentation of temporal bone structures from patients' conventional computed tomography (CT) data plays an important role in the image‐guided cochlear implant surgery. Existing convolutional neural network approaches have difficulties in segmenting such small tubular structures. Methods We propose a light‐weight three‐dimensional convolutional neural network referred to as W‐Net to achieve multiobjective segmentation of temporal bone structures including the cochlear labyrinth, ossicular chain and facial nerve from conventional temporal bone CT images. Data augmentation with morphological enhancement is proposed to increase the segmentation accuracy of small tubular structures. Evaluation against the state‐of‐the‐art methods is performed. Results Our method achieved mean Dice similarity coefficients (DSCs) of 0.90, 0.85 and 0.77 for the cochlear labyrinth, ossicular chain and facial nerve, respectively. These results were also close to the DSCs between human expert annotators (0.91, 0.91 and 0.72). Conclusions Our method achieves human‐level accuracy in the segmentation of the cochlear labyrinth, ossicular chain and facial nerve.

Fully automated segmentation in temporal bone CT with neural network: a preliminary assessment study

“…The proposed network model (W-Net) was described in detail in a previous study [ 30 ]. The W-net architecture in this study has no changes from the previously reported study [ 30 ]. Figure 1 shows the architecture of W-net.…”

“…Our previous study used data enhancement technology and 30 temporal bone CT volumes (15 left, 15 right) to develop the neural network model, which allowed the architecture to locate and render important structures on the left or right sides of the temporal bone CT [ 30 ]. We have used 24 volumes (12 left, 12 right) and 6 volumes (3 left, 3 right) as the training and validation set respectively, and trained the W-net network through five fold cross-validation [ 30 ]. It should be noted that the test set in this study were not included the training and validation data set in the previous study [ 30 ].…”

“…To address this gap in the field, we have proposed a deep-learning framework referred to as W-net architecture and performed preliminary training of the neural network to automatically segment critical structures in CT scans of normal adult conventional temporal bone [ 30 ]. To assess the feasibility of the model’s clinical application, we must further test the generalization ability of W-net.…”

Bayesian logistic shape model inference: Application to cochlear image segmentation

“…The former approaches are mostly based on cochlear shape fitting based on template image registration (Baker and Barnes, 2005), parametric shape model (Baker, 2008). The supervised methods are based on statistical deformation models (Ruiz Pujadas et al, 2018) and deep learning (Lv et al, 2021;Raabid et al, 2021;Heutink et al, 2020).…”

“…Deep Learning) is an effective way to perform image segmentation or processing in many cases. Specifically, in inner ear CT imaging analysis, many works achieved impressive results (Lv et al, 2021;Raabid et al, 2021;Heutink et al, 2020;Wang et al, 2019;Li et al, 2021;Alshazly et al, 2019;Zhang et al, 2019;Wang et al, 2020b). However, supervised learning methods have also many limitations.…”