A systematic review of automated temporal bone segmentation methods

Abstract

The temporal bone is a complex anatomical structure crucial for otologic and neurotologic procedures. Accurate segmentation of the temporal bone from computed tomography (CT) and magnetic resonance imaging (MRI) is essential for surgical planning, pathology assessment, and computational modeling. Manual segmentation is time-consuming and subject to inter-observer variability, necessitating the development of automated methods. This systematic review aims to analyze the current state of automated temporal bone segmentation techniques and their performance. A comprehensive search was conducted across PubMed, IEEE Xplore for articles published from 2004 to 2024. A total of 419 articles were reviewed, from which 34 were selected for this study. Among the identified studies, deep learning, particularly convolutional neural networks (CNNs) and U-Net variants, emerged as the dominant approach, consistently outperforming SSM and atlas-based methods. Deep learning models achieved the highest Dice Similarity Coefficient (DSC) and the lowest Hausdorff Distance (HD). Deep learning-based approaches improved automated temporal bone segmentation, with strong performance in segmenting larger structures such as the labyrinth, with Dice score over 0.86. However, the segmentation of smaller anatomical structures, such as stapes and chorda tympani, remains a challenge.