文献类型：期刊文献

英文题名：Uncertainty-Aware Upper Airway Segmentation for Non-Obese OSA

作者：Wu, Mingjuan[1]; Lu, Rihui[2]; Lin, Cong[3]; Zou, Lilan[3]; Qiu, Chenghao[3]; Fang, Yongjin[2]

机构：[1] Zhejiang University School of Medicine, Department of Health Management, Jinhua, China; [2] Zhejiang University School of Medicine, Department of Otolaryngology Head and Neck Surgery, Jinhua, China; [3] Guangdong Ocean University, School of Electronics and Information Engineering, Zhanjiang, China

年份：2026

起止页码：176

外文期刊名：2026 2nd International Conference on Unmanned Systems and Technology, UST 2026

收录：EI(收录号：20263721453604)

语种：英文

外文关键词：Diagnosis - Image segmentation - Medical imaging - Respiratory mechanics - Sleep research - Uncertainty analysis

外文摘要：Obstructive Sleep Apnea (OSA) remains a critical public health challenge, with non-obese patients being particularly difficult to diagnose due to complex anatomical abnormalities and dynamic soft-tissue collapse. In CT imaging, these factors manifest as blurred boundaries and high inter-patient variability, making manual annotation costly and inconsistent. Traditional fully supervised methods struggle to balance performance with the high demand for pixel-level labels. To bridge this gap, a label uncertainty-aware segmentation framework is proposed, which quantifies annotation noise through local standard deviation. By dynamically down-weighting high-uncertainty regions within the loss function, the model suppresses erroneous gradients and learns robust features. Validated on upper airway data with imprecise labels, this noise-robust framework achieves competitive performance using only low-cost coarse annotations. These results demonstrate significant robustness to noisy annotations and highlight the model's potential to alleviate the clinical burden of high-quality data acquisition. ? 2026 IEEE.