基于深度学习的舌骨上肌超声自动扫描线定位用于定量吞咽困难评估

文献信息

中文摘要

背景： 基于超声的舌骨上肌（SHM）运动评估已成为评估吞咽困难的一种有前景的方法。然而，定量分析仍受限于操作者依赖的扫描线放置和手动解读的变异性。本研究提出一个整合框架，结合数学建模、自动SHM分割、扫描线确定和基于深度学习的分析，用于客观的吞咽困难评估。方法：共分析了371张来自正常、轻度和重度吞咽困难患者的超声图像。临床相关的SHM衍生变量，包括厚度、位移、SHM差值（Δ = D - T）和总持续时间，被纳入数学解读框架。采用U-Net进行SHM分割和自动扫描线确定，并使用基于中位数的准确度、组内相关系数（ICC）、平均成对绝对差（MPAD）和总体标准差（SDpop）评估与专家选择的扫描线的一致性。此外，比较了五种深度学习架构（ResNet-101、Fast R-CNN、YOLO11、VGG-19和U-Net）用于吞咽困难严重程度分类。结果：所提出的框架与专家评估表现出高度一致性。U-Net在三个主要SHM组件上实现了平均交并比（IoU）为0.9599，表明接近专家的分割性能。自动扫描线确定与专家共识的总体一致性达到72.3%，而手动扫描线放置表现出相对较低的观察者间信度（ICC = 0.386）。在评估的分类模型中，Fast R-CNN表现最佳，F1分数为0.993，推理时间为3.44毫秒。Grad-CAM分析进一步证实，模型注意力集中在临床相关的SHM区域。结论：所提出的框架通过整合数学建模与AI驱动的分割和分类，为基于超声的吞咽困难评估提供了一种客观且可重复的方法。自动SHM定位和扫描线确定减少了操作者依赖性，同时保持专家级准确性。这些发现支持AI辅助超声作为标准化吞咽困难评估和实时决策支持的临床适用工具的潜力。

英文摘要

Ultrasound-based assessment of suprahyoid muscle (SHM) motion has emerged as a promising approach for evaluating dysphagia. However, quantitative analysis remains limited by operator-dependent scanline placement and variability in manual interpretation. This study proposes an integrated framework combining mathematical modeling, automated SHM segmentation, scanline determination, and deep learning-based analysis for objective dysphagia assessment. A total of 371 ultrasound images acquired from patients with normal, mild, and severe dysphagia were analyzed. Clinically relevant SHM-derived variables, including thickness, displacement, SHM Difference (Δ = D - T), and total duration, were incorporated into a mathematical interpretation framework. U-Net was employed for SHM segmentation and automatic scanline determination, and agreement with expert-selected scanlines was evaluated using median-based accuracy, intraclass correlation coefficient (ICC), mean pairwise absolute difference (MPAD), and population standard deviation (SDpop). In addition, five deep learning architectures (ResNet-101, Fast R-CNN, YOLO11, VGG-19, and U-Net) were compared for dysphagia severity classification. The proposed framework demonstrated high agreement with expert assessments. U-Net achieved an average intersection-over-union (IoU) of 0.9599 across the three major SHM components, indicating near-expert segmentation performance. Automated scanline determination achieved an overall agreement of 72.3% with expert consensus, while manual scanline placement exhibited relatively low interobserver reliability (ICC = 0.386). Among the evaluated classification models, Fast R-CNN achieved the highest performance with an F1-score of 0.993 and an inference time of 3.44 ms. Grad-CAM analysis further confirmed that model attention was concentrated on clinically relevant SHM regions. The proposed framework provides an objective and reproducible approach for ultrasound-based dysphagia assessment by integrating mathematical modeling with AI-driven segmentation and classification. Automated SHM localization and scanline determination reduce operator dependency while maintaining expert-level accuracy. These findings support the potential of AI-assisted ultrasound as a clinically applicable tool for standardized dysphagia evaluation and real-time decision support.