Deep learning networks for quantitative evaluation of organic voice disorders and their treatment

目前的诊断和治疗程序的有机语音障碍仍然占主导地位的主观/客观评价的声学信号和主观评价的振动特性的声带。这种主观评价主要是由于涉及复杂的物理过程以及声带和喉部的有限可及性，这阻碍了定量测量和评价。然而，机器学习和数值建模领域的最新科学进展已经证明了为许多疾病（包括语音障碍）的定量诊断和治疗提供支持的潜力。在本提案中，我们将利用最先进的神经元网络分类方法，结合数值建模和基于多模态的数据，通过提供治疗前、治疗期间和治疗后严重程度状态的定量信息，支持器质性嗓音障碍的临床治疗。第一个创新方面和工作包是开发一个改进的集中质量模型（6 MM+），具有逼真的Lattice-Boltzmann气流求解器，以模拟声带振动，利用现代图形处理单元（GPU）的优势。(2)我们将利用深度神经网络优化6 MM+动态，以实现声带振动的内窥镜高速记录。这将产生由新的6 MM+模型提供的生物力学喉部参数（即局部质量、刚度、声带碰撞力和声门下压力）。(3)将估计的生物力学参数、从内窥镜高速记录中提取的代表声带和声门动力学的参数、从声学信号和患者特定数据计算的参数组合成多模态数据集。然后，我们将再次利用这个多模态数据集的深度神经网络来量化手术治疗前、手术期间和手术后的器质性嗓音障碍的严重程度。(4)我们将利用这个多模态数据集内的特征重要性分析（自适应提升）来识别代表器质性语音障碍的重要参数。(5)为了实现我们的目标，我们将对90名声带麻痹、肌肉萎缩或声带息肉患者和一个对照组（60名受试者）进行研究，以训练和验证VITALITy系统。临床医生将协助创建VITALITy，以实现未来的理想临床应用。因此，VITALITy系统将有助于为器质性嗓音障碍的诊断、治疗进展和结果的定量评估提供所需的支持。该提案将通过结合最先进的机器学习、数值建模和多模态数据，为器质性嗓音障碍严重程度的定量评估设定新的标准。

Current diagnostic and treatment routines for organic voice disorders are still dominated by subjective/objective evaluation of the acoustical signal and subjective evaluation of the vibrational characteristics of the vocal folds. This subjective evaluation is mostly due to the complex physical processes involved and the limited accessibility of the vocal folds and larynx, which hinder quantitative measurements and evaluations. However, recent scientific advances in the areas of Machine Learning and numerical modeling have demonstrated the potential to contribute to quantitative diagnosis and treatment support for many diseases, including for voice disorders. In this proposal we will exploit stat-of-the-art neuronal network classification methods in combination with numerical modelling and multi-modal based data to support clinical treatment of organic voice disorders by providing quantitative information on severity status, pre- during and after treatment. The first innovative aspect and working package is the development of an improved lumped mass model (6MM+) with a realistic Lattice-Boltzmann airflow solver to simulate vocal fold vibrations, utilizing the advantages of modern Graphics Processing Units (GPUs). (2) We will utilize a deep neural network to optimize the 6MM+ dynamics towards endoscopic high-speed recordings of the vocal fold vibrations. This will yield biomechanical laryngeal parameters (i.e. local masses, stiffness, collision forces of vocal folds and subglottal pressure) provided by the new 6MM+ model. (3) The estimated biomechanical parameters, parameters representing vocal fold and glottis dynamics extracted from endoscopic high-speed recordings, parameters computed from the acoustic signal and patient specific data will be combined to a multi-modal data set. Then, we will again utilize a deep neural network on this multi-modal data set to quantify the severity of the organic voice disorder pre-, during and past surgical treatment. (4) We will utilize feature importance analysis (Adaptive Boosting) within this multi-modal data set to identify the important parameters representing organic voice disorders. (5) The entire work-flow will be integrated in the VoIce Treatment AnaLysIs Tool (VITALITy) software and made available for other researchers.To achieve our goals, we will conduct a study on 90 patients suffering from vocal fold paresis, muscle atrophy or vocal fold polyps and a control group (60 subjects) to train and validate the VITALITy system. Clinicians will assist in the creation of VITALITy to allow the aspired clinical application in future. Thereby, the VITALITy system will help to provide the desired support on quantitative evaluation for diagnostics, therapy progress and outcome of organic voice disorders.This proposal will set new standards in the quantitative evaluation of the severity of organic voice disorders by combining state-of-the-art machine learning, numerical modeling and multi-modal data.