- 學位論文
應用動態模擬與機器學習於早期圓錐角膜診斷
Dynamic Simulation and Machine Learning for Early Stage Keratoconus Detection
摘要
目前主要以角膜地形圖來診斷圓錐角膜，然而針對早期圓錐角膜患者，因角膜外觀尚未有明顯變化，診斷具有相當挑戰性。近年來，研究證實圓錐角膜首發於角膜內部膠原纖維結構的改變，進而引起不穩定的角膜生物機械性質。因此若能透過具備量測角膜生物力學性質能力的機器，如非接觸式眼壓計，獲得角膜動態反應與機械性質參數，將對早期圓錐角膜診斷具有極大的貢獻。 本論文主旨在於探討角膜於Corvis ST吹氣下的動態反應和圓錐角膜缺陷分布的關係。主要分為兩大部分，有限元素法與機器學習。首先透過有限元素法，建構圓錐角膜模型，模擬Corvis ST吹氣流程，分析設定缺陷參數對於角膜動態的影響。模擬結果發現，對稱型模態主要受缺陷位置(θ)、缺陷大小、缺陷機械性質與眼內壓的影響；而非對稱模態則受缺陷位置(θ、∅)、缺陷大小的影響。 除此之外，本論文提出兩種非監督式分類演算法，分別為基於形態與基於特徵的方法。藉由基於形態方法，建立診斷圓錐角膜的模態係數基準，發現對稱性模態與嚴重性、非對稱性模態與缺陷形心位置，皆具有高度相關性(皮爾森相關係數≥0.5)。在基於特徵的分類演算法中，分別使用三種不同的降維方法，其中以Isomap的效果最好，其對稱性模態與嚴重性的相關性達0.92，非對稱性模態與缺陷形心位置的相關性則達到0.75。
並列摘要
Keratoconus detection in early stage is challenging because of lacking of apparent clinical signs. Therefore, interests in clinical devices that can measure corneal biomechanical properties has largely increased these years. Since the initiating event in keratoconus is the biomechanical instability that caused by subtle changes to the corneal microscopic structure. Corvis ST is a non-contact tonometer that monitors corneal behavior under air puff with an ultra-high speed Scheimpflug camera and produces a set of biomechanical parameters. Assessing corneal dynamic behavior and biomechanical properties gives potential to aid keratoconus diagnosis in early stage when topography seems normal. The main purpose of this thesis is to investigate the relation between corneal dynamic behavior and keratoconic features. Finite element simulation of Corvis ST is established to study the influence of keratoconic features on Legendre modal parameters. Results showing θ, size, stiffness and IOP are pointed as important features in controlling symmetric mode, while asymmetric mode is determined by θ,∅ and the size of weak region in keratoconus. In addition, two unsupervised clustering algorithms, shape-based method and feature-based method with three different dimensionality reductions (PCA, NMF and Isomap) are proposed to construct keratoconic benchmarks. With shape-based method, strong correlations (Pearson product moment correlation (PPMC)≥0.5) are observed both in symmetric mode versus severity and asymmetric mode versus weak centroid of keratoconus. As for feature-based method, clustering using Isomap with six Legendre modes as input shows best result among all dimensionality reduction methods. PPMC between symmetric mode and normality achieves 0.92, while PPMC between asymmetric mode and weak centroid value attains 0.75. With our keratoconic benchmarks and clustering methods, information such as severity and the weak region of keratoconus are able to be obtained with Corvis ST.