| 摘要: | 近期，慢性肝病已成為對健康產生嚴重影響的疾病。在肝纖維化的早期，有效的 治療和適當的飲食管理可能有助於康復。然而，目前常用的無侵入性診斷方法，如血 液檢查和腹部超聲，對於肝纖維化的早期階段檢測並不十分靈敏。因此，在這項研究 中，我們提出了使用機器學習(ML)技術，透過特徵提取的方式，以預測肝纖維化階 段的方法。此研究結果有兩個部分，分別為動物的小鼠實驗與人類臨床數據。
小鼠實驗部分，分為健康(A0 )、輕度(A1 )、中度(A2 )或重度(A3 )四類。具 體來說，我們採用支持向量機作為分類器。使用兩種類型的特徵，即來自小鼠磁共振 (MR)成像的歐拉數(EN)與肝纖維化相關的估計孔隙率，用於訓練分類器。這兩種 特徵提取的基本想法源自於肝臟成像的幾何和拓撲特性。在參數調整之後，最終，在
模型比較方面，使用兩種特徵分別為 binary(Acc=63.2%)和 grayscale(Acc=64.5%) 影像給 CNN 模型訓練，和在 SVM 模型上，不同特徵作為訓練，分別為單獨的估計孔 隙率(Acc=93.3%)、單獨的歐拉特徵數(Acc=74.6%)和全部特徵(Acc=90.9%)。使用單獨的估計孔隙率特徵的模型在四類別的綜合表現上優於其他模型，特別是在輕度(A1 )類別中表現卓越。
進一步地，我們轉向人類肝纖維化分類，使用了 62 位病人的臨床數據，包括血 液檢查和 DCE-MRI 訊號曲線。為了增加準確性，我們引入了模擬肝臟的新特徵，如 孔隙率、擴散率、肝門靜脈和肝動脈的流速。通過結合這些特徵，我們使用 KNN 和 Naive Bayes 模型在 F0-3 vs F4-6 和 F0-5 vs F6 分類中取得了優異的結果，綜合三類精 確度保持在 69.4%。這項研究強調了模擬肝臟訊號濃度模型在肝纖維化評估中的潛在價 值，同時提供了生物醫學研究者更深入的理解和新的研究方向。;In recent years, chronic liver disease has emerged as a condition significantly impacting health. Effective treatment and proper dietary management in the early stages of liver fibrosis may contribute to recovery. However, non-invasive diagnostic methods such as blood tests and abdominal ultrasound are not highly sensitive for detecting early stages of liver fibrosis. Therefore, in this study, we propose using machine learning (ML) techniques, employing feature extraction, to predict the stages of liver fibrosis. The study comprises experiments on mice and clinical data from human subjects.
In the mouse experiment section, mice were categorized into healthy (A0), mild (A1), moderate (A2), or severe (A3) stages. Specifically, we utilized support vector machines as classifiers, using two types of features: Euler numbers (EN) from mouse magnetic resonance (MR) imaging and estimated porosity related to liver fibrosis. The basic idea behind these feature extractions stems from the geometric and topological properties of liver imaging. After parameter tuning, the final model comparisons showed that using the two features separately for binary (Acc=63.2%) and grayscale (Acc=64.5%) images for training CNN models, as well as on SVM models using different features—solely esti- mated porosity (Acc=93.3%), solely Euler characteristic numbers (Acc=74.6%), and all features combined (Acc=90.9%). The model using solely estimated porosity as a feature outperformed other models’ overall performance across the four categories, particularly excelling in the mild (A1) category.
Furthermore, we turned to human liver fibrosis classification, utilizing clinical data from 62 patients, including blood tests and dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) signal curves. To enhance accuracy, we introduced new features simulating liver properties, such as porosity, diffusion rate, and flow speeds of the portal vein and hepatic artery. Combining these features, we used KNN and Naive Bayes models to achieve excellent results in F0-3 vs. F4-6 and F0-5 vs. F6 classifications, with an overall three-class accuracy of 69.4%. This study underscores the potential value of simulating liver signal concentration models in liver fibrosis assessment, providing biomedical researchers with a deeper understanding and new avenues for research. |
