角膜噴氣試驗分析與圓錐角膜歷程追蹤

Abstract

圓錐角膜是最常見的原發性角膜擴張症，特徵是局部角膜變薄，進而導致角膜產生突出，造成不規則散光而影響視覺。疾病的進展表現無法使用眼鏡來補償，故會導致患者持續替換眼鏡。傳統檢測中，除非進行角膜地形圖和角膜剖面檢查，否則難以被發現。傳統檢測儀器大多屬於靜態角膜特徵的觀察，若要進行動態特徵觀察，則可使用眼壓計。它的原理是利用吹氣使角膜產生變形，本研究藉此方法來發現區分圓錐角膜。透過蒐集Corvis®ST 眼壓計吹氣檢測數據，包含散光、頓挫型圓錐角膜、圓錐角膜患者的角膜數據以及正常人的角膜數據，配合角膜數學理論進行數據的模態展開，並利用動力學理論萃取出疾病的特徵參數，使用疾病辨識法則制定分類規則，並進行t-SNE數據視覺化與疾病分類。針對長時間在臨床上有追蹤的患者進行追蹤，將挑選出的患者數據進行正規化，再經由時間定義來尋找出疾病進展趨勢。進行完t-SNE視覺化處理後，可經由人眼辨識正常人與疾病之間的分類以及散光與圓錐角膜分類、頓挫型圓錐角膜與圓錐角膜分類，藉此增加機器學習在分類後的可信度。疾病分類使用決策樹、k-近鄰、支持向量機、邏輯斯回歸與人工神經網路，進行正常人與疾病之間分類比較以及疾病倆倆之間分類比較。結果發現以人工神經網路與決策樹在正常人與疾病、散光與圓錐角膜、頓挫型圓錐角膜與圓錐角膜中分類具有良好的準確度，其中人工神經網路在區別正常人與疾病患者時，平均準確度最高可達97.72%，而將正常人與三種疾病同時分類時，左眼可達74.37%，右眼可達78.63%。利用決策樹還可以找出疾病分類中特別優秀的3個特徵。最後針對疾病追蹤，可統計出三種不同的趨勢，分別為延展、來回以及突出，其中，來回為圓錐角膜中最常見的趨勢。本研究使用了自行建立的特徵配合機器學習的方法執行疾病分類，並提出合適追蹤的特徵適合用於未來病況之追蹤。

Keratoconus is the most common form of primary corneal ectasia, characterized by localized thinning of the cornea. It leads to protrusion and irregular astigmatism, affecting vision. Traditional diagnostic instruments rely on static observations of corneal features, while dynamic observation with an air-puff tonometer is currently the advanced biomechanical method. Air-puff deforms the cornea by blowing air, and it could be the method to detect keratoconus at an early stage. In our study, we collected data from Corvis®ST air-puff tonometry, including corneal data from patients with astigmatism, keratoconus, and normal individuals. Through mathematical theory, Legendre decomposition, and other machine learning techniques, we extracted features of the keratoconus. By applying disease recognition algorithms, mode decomposition, t-SNE data visualization, and disease classification, we aimed to identify disease patterns. Then we performed normalization for long-term patient tracking, seeking disease progression trends based on time definitions. The results showed that artificial neural networks and decision trees achieved high accuracy in distinguishing between normal individuals and disease, astigmatism and keratoconus, and forme fruste keratoconus and keratoconus. The artificial neural network achieved an average accuracy of up to 97.72% in distinguishing between normal individuals and patients, and 74.37% for left eyes and 78.63% for right eyes in classifying the four disease types. Additionally, the decision tree identified three particularly excellent features for disease classification. Finally, for disease tracking, we identified three different trends: progressive, oscillatory, and protruding. Among them, the oscillatory trend was the most common in keratoconus. To conclude, this study utilized self-established features combined with machine learning methods for disease classification and proposed appropriate features for future disease tracking.