Objective analysis of functional based hoarseness by clinical high-speed endoscopy

以功能性发音困难为基础的声音嘶哑在日常临床中相当常见。据统计，功能性发音障碍可能占所有被诊断为发音障碍的50%。与易于识别的基于形态的发音障碍相比，功能性发音困难只在发声时，即在振动期间显示其临床图像。这些病理性振动模式从没有声门间隙的周期性轻微动态左右不对称，到高度动态的左右不对称，甚至是伴有声门间隙的不规则振动。为了可视化声带动力学，我们进行了内窥镜数字高速视频成像(HSV)并同步记录发出的声学语音信号。到目前为止，基于HSV的诊断仍然广泛地主观执行，因此很大程度上依赖于医生的经验。因此，原因仍然是缺乏自动化的临床适用的图像处理算法，因此缺乏普遍接受的客观HSV参数。在目前的项目中，到目前为止，我们开发了一个软件工具，允许自动提取客观的HSV参数；即足够快、有效和健壮。该软件已经被7个国家的27个研究小组使用。我们发现，许多实用的HSV参数不适合描述功能性发音困难，也有许多参数根本不适合临床应用。应用最先进的机器学习算法，我们分离了不同类型的功能性发音障碍，并将这些发音障碍与健康的语音产生分开。然而，目前的准确性还不足以用于临床应用，因为分类任务是针对不同的传感器数据单独执行的：声学、HSV成像和临床评估工具。我们将在下一个项目阶段克服这一缺点：因此，这个项目阶段的中心目标是通过应用机器学习技术对所有多传感器(声学、HSV成像、进一步的临床评估工具)数据进行综合分析，以(1)客观地对声音嘶哑进行分级；(2)确定与年龄有关的参数；(3)客观地评估和量化治疗进展；(4)将开发的机器学习算法实施在软件工具中，然后可以被其他研究和临床小组使用，最终将这些机器学习方法转移到临床应用；即，基于计算机的临床状态的定量和可视呈现，以评估功能性发音困难的临床图像和治疗进展。

Hoarseness based on functional dysphonia is rather common in daily clinical routine. According to statistics, functional dysphonias may state up to 50% of all diagnosed disordered voices. In contrast to easily to recognize morphological based voice disorders, functional dysphonias show their clinical picture only during phonation, i.e. during vibration. These pathological vibration patterns range from periodic slightly dynamic left-right asymmetry without glottal gap to highly dynamic left-right asymmetry or even irregular vibrations combined with a glottal gap. To visualize vocal fold dynamics, we perform endoscopic digital high-speed video imaging (HSV) and synchronously record the emitted acoustic voice signal. So far, HSV based diagnostics is still widely performed subjectively and therefore depends strongly on the experience of the medical doctor. Reasons therefore are still missing automated clinically applicable image processing algorithms and hence the lack of commonly accepted objective HSV parameters. Within the current project, so far, we developed a software tool allowing for automatic extraction of objective HSV parameters; i.e. sufficient fast, valid and robust. This software is already used by 27 research groups in 7 countries. We showed that many applied HSV parameters are not suitable for characterizing functional dysphonia and that many parameters are also not suitable for clinical application at all. Applying state-of-the-art machine learning algorithms we separated different kinds of functional dysphonia and separated these also from healthy voice production. However, the current accuracy is not yet sufficient for clinical use, since the classification tasks were performed separately for the different sensor data: acoustics, HSV imaging and clinical assessment tools. We will overcome this shortcoming in the next project phase:Hence, the central goal in this project phase is a combined analysis of all multi-sensor (acoustics, HSV imaging, further clinical assessment tools) data by applying machine learning techniques to (1) objectively grade hoarseness; (2) determine age dependent parameters; (3) objectively assess and quantify treatment progress; (4) implement the developed machine learning algorithms in a software tool that then can be used by other research and clinical groups to finally transfer these machine learning methods to clinical application; i.e. a computer based quantitative and visual presentation of the clinical status for assessment of the clinical picture of functional dysphonia and treatment progress.