角膜神经图像中的自动检测与疾病分析技术研究

研究表明，多种眼科疾病和慢性疾病都会损伤角膜神经网络且引起其病理形态的变化。因此，角膜神经的几何和拓扑自动量化分析，对多种眼科及全身性疾病的筛查、诊断和评估至关重要。现有方法存在多方面挑战：图像对比度低，光照不均衡，导致图片可读性差；细微神经连通性差、拓扑信息缺失，导致难以对神经形态进行自动提取和个体化分析；基于人工标注的形态分析存在主观性强和效率低的特点。如何在理论模型和关键技术上实现神经结构精准提取和拓扑信息获取，是突破角膜神经形态分析局限的关键科学问题。本项目以上这些问题，拟开展以下研究：1、以图像增强为手段，研究基于注意力机制机制的神经结构自动分割方法，提升细微和非连续神经的分割准确性；2、研究优势集聚类方法，将神经结构的连接关系转为聚类问题，实现角膜神经网络拓扑结构建立； 3、研究角膜神经弯曲度语义特征提取方法，现对角膜神经弯曲度自动评估，增强自动评估模型的精度和可解释性。

Precise characterization and analysis of corneal nerve fibre are of great importance in facilitating examination and diagnosis of many eye-related diseases. Recent clinical researches suggest that the fibre tortuosity changes might associate with a number of diseases. This calls for precise fibres tracing and their tortuosity description. Existing automated methods are often problematic due to poor contrast, imaging noise and fragmentation of fibers caused by imperfect focusing. To address these problems, in this proposal, we propose a fully automated framework for corneal nerve analysis, consisting of image enhancement, fibre detection, topology estimation, and tortuosity level classification. We first use the Retinex theory to correct the imaging illumination, then introduce a curvilinear structure attention network to extract the nerve fibre, and the dominant set clustering is utilized to estimate the topology of this fires, finally the tortuosity level of each image is obtained by using fuzzy rough model. This framework will become a powerful tool for management of a wide spectrum of nerve-related diseases.