UADSN: Uncertainty-Aware Dual-Stream Network for Facial Nerve Segmentation

Guanghao Zhu, Lin Liu, Jing Zhang, Xiaohui Du, Ruqian Hao, Juanxiu Liu

Abstract

Facial nerve segmentation is crucial for preoperative path planning in cochlear implantation surgery. Recently, researchers have proposed some segmentation methods, such as atlas-based and deep learning-based methods. However, since the facial nerve is a tubular organ with a diameter of only 1.0-1.5mm, it is challenging to locate and segment the facial nerve in CT scans. In this work, we propose an uncertainty-aware dual-stream network (UADSN). UADSN consists of a 2D segmentation stream and a 3D segmentation stream. Predictions from two streams are used to identify uncertain regions, and a consistency loss is employed to supervise the segmentation of these regions. In addition, we introduce channel squeeze & spatial excitation modules into the skip connections of U-shaped networks to extract meaningful spatial information. In order to consider topology preservation, a clDice loss is introduced into the supervised loss function. Experimental results on the facial nerve dataset demonstrate the effectiveness of UADSN and our submodules.

Introduction

In cochlear implantation surgery, surgeons must navigate the temporal bone to reach the cochlea. A critical risk during this procedure is damaging the facial nerve, a tiny tubular structure with a diameter of only 1.0–1.5mm. Accidental injury can lead to permanent facial paralysis. While high-resolution CT scans provide the necessary anatomical data, the nerve's small size and blurry boundaries make manual segmentation extremely difficult for radiologists.

Traditional atlas-based methods often fail to account for individual anatomical variations. Meanwhile, standard deep learning models frequently struggle with disconnected segmentation results or overfitting due to the extreme class imbalance between the tiny nerve and the surrounding tissue. To address these challenges, we present the Uncertainty-Aware Dual-Stream Network (UADSN).

Methodology

The UADSN architecture is designed to leverage the complementary strengths of 2D and 3D processing while ensuring topological integrity.

Dual-Stream Architecture

The model processes CT data through two parallel paths: a 3D segmentation stream to capture volumetric context and a 2D segmentation stream that utilizes a larger receptive field with fewer parameters, reducing the risk of overfitting on small datasets.

Overview of the UADSN architecture, featuring parallel 2D and 3D segmentation streams and uncertainty-aware consistency loss.

Uncertainty-Aware Consistency Loss

By comparing predictions from the 2D and 3D streams, the network identifies "uncertain regions" where the streams disagree. A consistency loss is then applied to these specific areas, forcing the model to refine its predictions at difficult boundary locations.

Topology Preservation and Attention

To maintain the connectivity of the tubular nerve, we integrate clDice (centerline Dice) loss. Unlike standard Dice loss, which only measures overlap, clDice ensures the "skeleton" of the nerve remains continuous. Additionally, spatial Squeeze & Excitation (sSE) modules are embedded in the skip connections to help the network focus on relevant spatial features while suppressing background noise.

Experiments and Results

We evaluated UADSN on a self-built dataset of 28 temporal bone CT images with expert annotations. The data was split into 23 training cases and 5 evaluation cases, using cropped 64x64x32 blocks for training.

UADSN achieved a Dice Similarity Coefficient (DSC) of 79.79%, outperforming 3D U-Net, Attention U-Net, and the Transformer-based UNETR. Notably, UNETR performed significantly worse (63.16% DSC), likely due to the limited size of the training dataset. UADSN also achieved the lowest Average Symmetric Surface Distance (ASSD) of 0.0952 mm, indicating high boundary precision.

Qualitative comparison of segmentation results. UADSN (Ours) shows fewer false positives (yellow arrows) compared to other state-of-the-art models.

Qualitative analysis demonstrates that UADSN is less prone to false positives and maintains better connectivity than competing methods, even without post-processing.

Conclusion

The UADSN framework demonstrates that combining 2D and 3D features with uncertainty-aware supervision and topology-preserving loss functions significantly improves the segmentation of tiny tubular structures like the facial nerve. While the current results are promising, future work will focus on validating the model on larger, multi-center datasets and exploring more sophisticated uncertainty quantification metrics for clinical deployment.

Summary updated

Abstract

Facial nerve segmentation is crucial for preoperative path planning in cochlear implantation surgery. Recently, researchers have proposed some segmentation methods, such as atlas-based and deep learning-based methods. However, since the facial nerve is a tubular organ with a diameter of only 1.0-1.5mm, it is challenging to locate and segment the facial nerve in CT scans. In this work, we propose an uncertainty-aware dual-stream network (UADSN). UADSN consists of a 2D segmentation stream and a 3D segmentation stream. Predictions from two streams are used to identify uncertain regions, and a consistency loss is employed to supervise the segmentation of these regions. In addition, we introduce channel squeeze & spatial excitation modules into the skip connections of U-shaped networks to extract meaningful spatial information. In order to consider topology-preservation, a clDice loss is introduced into the supervised loss function. Experimental results on the facial nerve dataset demonstrate the effectiveness of UADSN and our submodules.
BibTeX
@misc{zhu2024uadsn,
  author = {Guanghao Zhu and Lin Liu and Jing Zhang and Xiaohui Du and Ruqian Hao and Juanxiu Liu},
  title = {{UADSN: Uncertainty-Aware Dual-Stream Network for Facial Nerve Segmentation}},
  year = {2024},
  eprint = {2407.00297},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2407.00297},
}
Figure