共轭贝叶斯证据学习用于不确定性感知的鼻咽癌分割

文献信息

中文摘要

鼻咽癌（NPC）大体肿瘤体积（GTV）的准确自动分割对于放疗计划至关重要，然而传统的确定性模型无法充分表征磁共振成像（MRI）上模糊肿瘤边界固有的不确定性。现有的用于不确定性量化的证据深度学习（EDL）方法通常依赖固定先验，理论上容易出现过证据饱和，从而限制了不确定性估计的可靠性。为了解决这些问题，我们提出了一种共轭贝叶斯证据分割（CoBESeg）框架，用于不确定性感知的NPC GTV分割。为了缓解传统EDL中的证据饱和，引入了一个分层方差证据网络，通过共轭贝叶斯更新明确地将解剖先验知识与图像驱动的证据相结合。为了提高模糊边界区域的特征判别能力，提出了一种模糊原型对比学习策略，以增强证据表示的可判别性。此外，CoBESeg结合了一种特征距离感知的证据校准策略，在测试时动态校准证据强度。在三个多中心NPC MRI数据集上的实验表明，与最先进的方法相比，CoBESeg实现了更优的分割精度和更可靠的不确定性估计，同时支持临床偏移下的风险过滤。代码可在 https://github.com/zhangchi73/hvenfpcl 获取。

英文摘要

Accurate automatic segmentation of the gross tumor volume (GTV) in nasopharyngeal carcinoma (NPC) is critical for radiotherapy planning, yet conventional deterministic models do not adequately characterize the inherent uncertainty associated with ambiguous tumor boundaries on magnetic resonance imaging (MRI). Existing evidential deep learning (EDL) methods for uncertainty quantification typically rely on fixed priors and are theoretically prone to evidence saturation, thereby limiting the reliability of uncertainty estimation. To address these issues, we propose a conjugate Bayesian evidential segmentation (CoBESeg) framework for uncertainty-aware NPC GTV segmentation. To alleviate evidence saturation in conventional EDL, a hierarchical variance evidence network is introduced to explicitly integrate anatomical prior knowledge with image-driven evidence through conjugate Bayesian updating. To improve feature discrimination in ambiguous boundary regions, a fuzzy prototype contrastive learning strategy is proposed to enhance the discriminability of evidential representations. Furthermore, CoBESeg incorporates a feature distance-aware evidential calibration strategy to dynamically calibrate evidence strength at test time. Experiments on three multicenter NPC MRI datasets demonstrate that CoBESeg achieves superior segmentation accuracy and more reliable uncertainty estimation compared with state-of-the-art methods, while also supporting risk filtering under clinical shift. The code is available at https://github.com/zhangchi73/hvenfpcl.