Tumor Segmentation and Tumor Probability Analysis with Convolutional Neural Networks: Novel Strategies for Individualized Radiation Therapy of Head&Neck Cancer

在放射治疗计划中，手动分割肿瘤病灶是一项繁琐且依赖于用户的任务。自动分割算法已经被提出作为一种选择，其中卷积神经网络(CNN)在有足够大量的独立输入ImageData的情况下表现出最佳的分割性能。在头颈部肿瘤的放射治疗计划中，MRI、PET和CT的多对比度和多模式成像通常每次检查最多有9种不同的对比度。这些图像被用来描绘肿瘤边缘，并识别肿瘤中需要更高辐射剂量才能进行有效治疗的缺氧亚区。在这个项目中，CNN将根据头颈部肿瘤治疗的前瞻性临床试验的现有图像数据进行训练。由于CNN的内部参数的最佳结构或最佳设置都不是已知的，将使用现有的辐射计划作为基本事实来训练和比较多个CNN。此外，将开发预处理方法来校正系统图像的异质性(例如来自不同场强的MRI系统)，以便研究中可以包括更多的数据。为此，将使用已知的MRI信号方程来单独修改每种组织类型的信号强度。多通道、多对比度图像的获取非常耗时，为了识别那些不能提高分割性能的输入图像数据，采用留一策略对具有不同输入数据的CNN进行训练，并将CNN与完整的CNN实现进行比较。作为这种优化的结果，成像协议将被优化，并将实现具有最佳性能的CNN。在得到的CNN配置下，将执行自动的肿瘤和淋巴结分割，并与一组来自组织学的已建立的生物标记物相关联，以评估CNN衍生的成像特征是否可以预测单个肿瘤生物学。该项目的结果将是一个用于头颈部肿瘤自动分割的优化的CNN，它还将提供有关缩短诊断成像方案的方法和与缺氧相关的肿瘤异质性的额外放射组学信息，这可能对减少肿瘤复发的新策略产生直接影响。

The manual segmentation of tumor lesions is a tedious and user-dependent task in radiation therapy planning. Automatic segmentation algorithms have been proposed as an alternative, among which convolutional neural networks (CNN) exhibit the best segmentation performance if a sufficiently large amount of independent input imagedata is available. In radiation therapy planning of head&neck tumors, multi-contrast and multimodality imaging with MRI, PET and CT is often performed with up to 9 different contrasts per exam. These images are used to delineate the tumor margins and to identify hypoxic sub-regions in the tumor which require higher radiation dosesfor an efficient treatment. In this project CNNs will be trained on existing image data of a prospective clinical trial for head&neck tumor treatment. As neither the optimal architecture nor the best setting for the internal parameters is known for CNNs, multiple CNNs will be trained and compared using existing radiation plans as ground truth.In addition, pre-processing methods will be developed to correct for systematic image heterogeneity (originating for example from MRI systems with different field strength) so that more data can be included in the study. For this, the known MRI signal equations will be used to modify the signal intensities for each tissue type individually. It is time-consuming to acquire multimodality and multicontrast images.To identify those input image data that do not improve the segmentation performance, CNNs with different input data will be trained using a leave-one-out strategy, and the CNNs will be compared with a full CNN implementation. As a result of this optimization, the imaging protocol will be optimized and a CNN with optimal performance will be realized. With the resulting CNN configuration automatic tumor and lymph node segmentation will be performed, and correlated with a panel of established biological markers from histology to assess whether CNN-derived imaging features can predict individual tumor biology. The result of this project will be an optimized CNN for automaticsegmentation of head&neck tumors which will also provide additional radiomics information about ways to shorten the diagnostic imaging protocols and about hypoxia-related tumor heterogeneity, which might have direct consequences for new strategies to minimize tumor recurrence.