建立以機器學習為基礎之超音波多特徵脂肪肝定量技術

Abstract

脂肪肝是一種過多脂肪堆積於肝臟之疾病，若不及早透過健康飲食和運動改善，就有可能演變為肝硬化和肝癌等晚期肝臟疾病。在過去脂肪肝診斷方法中，病理學切片被視為黃金標準，但由於其侵入式所帶來的副作用和爭議，逐漸被非侵入式的醫療影像診斷所取代，而考慮到價格、安全和方便性等因素，超音波是最適合的診斷工具。 傳統的超音波參數上存在種種限制而無法適用於大多數的狀況，於是本研究從原始訊號中萃取出三種不同物理意義的超音波組織特性參數，來幫助判斷脂肪肝，包含集成逆散射(IB, 逆散射訊號強度的度量)、希爾伯特-黃轉換的Q因子(Q factor, 用於觀察頻率衰減的新參數)、均質性因子(HF, 量化脂肪均勻度的新參數)。 但單一參數仍然有其物理意義上的限制，因此本研究使用機器學習中的支持向量機的三個核函數作為演算法來結合上述三個參數（特徵），試圖以結合多特徵來突破單一特徵在其物理意義上的限制。並以A組(111筆)和B組(74筆)病例分別作為機器學習中的訓練和測試資料，10%脂肪變性來判斷是否為顯著脂肪肝。 結果表明萃取出的參數在各自表現上也有不錯的判斷脂肪肝之能力，而除了敏感度以外的所有診斷評判參數上都能透過多特徵結合而有所提升，且區分正常和脂肪肝患者之準確率達到86.49%且ROC曲線下面積達到0.8929，並找出各自適用於輔助懷疑和排除患病的兩種特徵組合。本研究提供了一種通用性高、計算複雜度較低且準確率高的脂肪肝判斷方法，在脂肪肝的輔助診斷上有發展潛力跟相當好的臨床應用價值。

Fatty liver is a disease which excess fat accumulates in the liver. If it is not improved through a healthy diet and exercise as early as possible, it may become terminal liver diseases such as cirrhosis and cancer. Pathology was considered as the gold standard method of diagnosing fatty liver in the past, but due to its invasive side effects and controversies, it was gradually replaced by non-invasive medical imaging diagnosis. Considering price, safety and convenience, ultrasound is the most suitable diagnostic tool. But there are many limitations in traditional ultrasonic parameters which make it not suitable in most circumstances. In view of this, we extracted three different physical characteristics of ultrasound tissue characteristics parameters from the original signal to help diagnosing the fatty liver, including the integrated backscatter (IB, a measure of backscatter signal intensity), the Q factor of the Hilbert-Huang transition (Q factor , a new parameter for observing frequency decay), and the homogeneity factor (HF, a new parameter for quantifying fat evenness). However, the single parameter still has its limitations in physical meaning; therefore we use the three kernel functions of the support vector machine in machine learning as an algorithm to combine the above three parameters (features), attempting to break the limitations by combining multiple features. Groups A (111 samples) and B (74 samples) are used as training and test data in machine learning respectively, and 10% steatosis is used to judge whether it was a significant fatty liver. The results show that the extracted parameters also have a good ability to judge fatty liver in their respective performances. Except for sensitivity, all diagnostic parameters can be improved by combining multiple features. The accuracy of identification between normal and fatty patients come to 86.49%, and the area under the ROC curve reach to 0.8929. Also, we find the two combinations of features that are suitable to assist in suspecting and excluding disease respectively. This study provides a method for judging fatty liver with high versatility, low computational complexity, and high accuracy with developing potential in the diagnosis of fatty liver and good clinical application value.