基于跨任务融合与时间上下文聚合模块的吞咽功能定量评估模型的构建与应用

文献信息

中文摘要

背景： 视频荧光吞咽研究（VFSS）是吞咽困难评估的临床金标准，但常规定量分析仍依赖逐帧手动标注和测量，导致时间成本高且评分者间一致性有限。现有的深度学习方法通常自动化孤立的子任务（例如关键点定位或时间阶段识别），而专门为完整定量分析流程设计的模型仍然罕见。我们提出了VFSS-QNet（VFSS定量分析网络），这是一种轻量级多任务模型，其编码器基于轻量级门控视觉Transformer（LGViT），并包含时间上下文聚合（TCA）和跨任务融合（CTF）模块。单个网络并行产生两个输出，即解剖关键点（C2、C4、舌骨以及上食管括约肌的内侧和外侧）和功能区域（食团、咽部和喉前庭）。对于预选的吞咽段，这些输出被转换为六条逐帧曲线，用于提取六个时间参数和五个运动学参数。使用回顾性数据集（43名受试者，35名训练和8名测试）的固定受试者级别划分，VFSS-QNet在关键点检测中实现了5.34像素的mLE和99.28%的mR2。对15个吞咽段的一致性分析显示，持续时间型和连续趋势驱动参数具有强一致性，而峰值敏感的幅度测量则不太稳定。VFSS-QNet进一步提供实时推理，速度为69.39帧/秒（FPS），参数量为4.22M，GPU内存使用量为0.35 GB。总体而言，该模型在准确性和计算效率之间取得了良好的平衡，对于减少手动工作负荷、支持标准化评估和吞咽功能个体化康复具有价值。

英文摘要

Videofluoroscopic Swallowing Study (VFSS) is the clinical gold standard for dysphagia assessment, yet routine quantitative analysis still relies on frame-by-frame manual annotation and measurement, resulting in a high time cost and limited inter-rater consistency. Existing deep-learning approaches typically automate isolated subtasks (e.g., keypoint localization or temporal stage recognition), whereas a model specifically designed for the full quantitative analysis pipeline remains rare. We propose the VFSS-QNet (VFSS quantitative analysis network), a lightweight multi-task model with an encoder based on a lightweight gated vision transformer (LGViT) and modules for temporal context aggregation (TCA) and cross-task fusion (CTF). A single network produces two outputs in parallel, namely anatomical keypoints (C2, C4, hyoid, and inner and outer sides of the upper esophageal sphincter) and functional regions (bolus, pharynx, and laryngeal vestibule). For preselected swallow segments, these outputs are converted into six frame-wise curves for extraction of six temporal and five kinematic parameters. Using a fixed subject-level split of a retrospective dataset of 43 subjects (35 training and eight testing), VFSS-QNet achieves an mLE of 5.34 px and an mR2 of 99.28% for keypoint detection. Consistency analyses on 15 swallow segments show strong agreement for duration-type and continuous trend-driven parameters, whereas peak-sensitive amplitude measures remained less stable. VFSS-QNet further provides real-time inference at 69.39 frames per second (FPS) with 4.22 M parameters and 0.35 GB of GPU memory usage. Overall, the model achieves a favorable balance among accuracy and computational efficiency, with value for reducing manual workload and supporting standardized assessment and individualized rehabilitation in swallowing function.