基于图像语义特征提取的角膜神经扭曲度智能评估方法研究

It has been proven in clinical research that the tortuosity of corneal nerve is associated with a variety of diseases and can assist screening and diagnosis of them. The manual assessment of tortuosity is subjective and inefficient. Sometimes, the assessment results of different doctors are inconsistent, which increases the difficulty of discovering relevant knowledge. This project extracts semantic features based on supervised learning from confocal microscope corneal nerve image with expert annotations, and establishes an automated tortuosity assessment model based on rule-learning. The main research contents include: the reliability-based annotation fusion method is proposed for training set enhancement; On the basis of the segmentation of neural fiber and the reconstruction of their topology, a medical image semantic feature extraction methods based on ordered weighted average and pre-aggregation functions are proposed; With exploiting the interpretability of rule-based models, the tortuosity assessment system based on fuzzy rule system and its hybrid model with deep neural network are constructed. Quantitative and qualitative analysis are employed to evaluate the accuracy and interpretability of the proposed system, respectively. It is expected that the research outputs of this project will improve the efficiency and intelligence of corneal nerve tortuosity assessment, support the diagnosis of related diseases, and provide new ideas and methods for medical image processing.

临床研究证明角膜神经的扭曲度变化与多种疾病相关,可为多种疾病的筛查和诊疗提供依据。人工扭曲度等级评估具有主观性强和效率低等缺点,不同医生的评估结果常常会不一致,增加了相关知识发现的难度。本项目拟针对共焦显微镜角膜神经图像的专家标注数据,研究基于监督学习的语义特征提取方法,并建立基于规则学习的自动扭曲度评估分类模型。主要研究内容包括:针对专家评估结果不一致的问题,提出基于可信度的标注融合方法用于训练集增强;在神经纤维分割和拓扑结构建立的基础上,提出基于有序加权平均和预聚合函数的医学图像语义特征提取方法;结合规则模型良好的可解释性,构建基于模糊规则模型及其与深度神经网络混合模型的扭曲度评估系统;分别采用定量和定性的方法对模型的精度及其可解释性进行评价。预期项目研究成果将提高角膜神经扭曲度评估的效率和智能化程度,为相关疾病的辅助诊断提供支持,为医学影像处理提供新研究思路。

临床研究证明角膜神经的扭曲度变化与多种疾病相关,可为多种疾病的筛查和诊疗提供依据。人工扭曲度等级评估具有主观性强和效率低等缺点,不同医生的评估结果常常会不一致,增加了相关知识发现的难度。本项目基于活体共焦显微镜角膜神经图像，针对神经扭曲度自动评估中的关键问题开展了研究,主要研究成果包括：1）实现了角膜神经图像数据采集、标注和图像增强，提出了一种基于模糊积分的多个深度学习模型集成框架，对重复标注数据中每个标注者进行建模，之后通过基于神经网络结构实现的Choquet积分来集成多个标注者的分割结果。2）提出了多种医学图像中线状结构的分割方法，并提出了基于优势集的线状结构拓扑重建方法。3）提出了一种不依赖于分割结果的角膜神经图像扭曲度计算方法，并分析了从分支级扭曲度到图像级扭曲度的映射，提出了基于聚合函数的图像级扭曲度特征提取方法，实现了扭曲度语义特征的提取和选择。4）提出了一种通过密度聚类构造模糊语义变量的无监督扭曲度自动评估方法，以及一种全自动深度学习方法用于角膜神经的图像级扭曲度分级，利用可解释性原则进一步提高分级精度。除此之外，还提出一种基于模糊规则的智能系统，通过数据学习获得准确和可解释的模糊规则库来支持临床决策。本项目研究成果提高了角膜神经扭曲度评估的效率和智能化程度,为相关疾病的辅助诊断提供支持,为医学影像处理提供新研究思路。