使用小波轉換估算電腦斷層掃描影像的肺分流比例

Abstract

肝癌是全球常見癌症位居第六名，佔癌症總量的5.6%，也是造成癌症相關死亡原因中的第三位。肝肺分流比例是篩選欲接受經肝動脈釔-90放射性栓塞治療患者的標準流程之一。肝肺分流比例測定的黃金標準是使用鎝-99m標記的白蛋白的肺血流灌注造影。然而，全球僅有五座反應爐提供鎝-99m，其中大部分老舊及衰退導致面臨鎝-99m短缺。本研究目的為利用計算的方法且不使用鎝-99m標記的白蛋白來估計肝肺分流比例值。 本研究納入77位臺北榮民總醫院肝癌病患之電腦斷層掃描影像，每位患者具有9個臨床特徵以及來自有顯影和無顯影電腦斷層掃描影像的小波特徵各380個。以新提出的SVR-LSF方法來從電腦斷層掃描影像中估計肝肺分流比例值。 SVR-LSF從有顯影電腦斷層掃描影像中挑選出57個特徵，並使用留一交叉驗證得到均方根誤差和相關係數分別為1.032和0.995；在無顯影電腦斷層掃描影像中挑選出62個特徵，並使用留一交叉驗證得到均方根誤差和相關係數分別為1.018和0.995。 研究結果顯示，本研究提出SVR-LSF方法能有效從眾多特徵中找出一組少數且誤差小的最佳化回歸方法，同時能建立回歸模型以及分類模型。加上臨床特徵集能增加有效篩選出能接受經肝動脈釔-90放射性栓塞治療的患者。

Hepatocellular carcinoma (HCC) is the sixth most common cancer, accounting for 5.6% of all cancer diagnoses, and it is also the third most common cause of cancer-related deaths worldwide. Lung shunt fraction (LSF) is one of the criteria to select patients who are willing to undergo yttrium-90 (Y-90) radioembolization of malignant liver tumors. The gold standard for LSF is to use technetium-99m macroaggregated albumin (Tc-99m MAA) lung perfusion imaging. However, there are just five reactors suppling Tc-99m worldwide and most of them age too rapidly to a widespread of Tc-99m shortage. This study aims to use a computational method to estimate LSF without using Tc-99m MAA. This study included 77 patients who underwent computed tomography (CT) in Taipei Veterans General Hospital. Each patient has 9 clinical features and 380 wavelet features from a contrast-enhanced CT and also another 380 wavelet features from a non-contrast-enhanced CT. This study proposed a computational method SVR-LSF to estimate LSF from a computed tomography (CT) image. SVR-LSF identified 57 features from contrast-enhanced CTs and achieved root mean square error (RMSE) and correlation coefficient (R) of 1.032 and 0.9947, respectively, with leave-one-out cross validation (LOOCV) method. SVR-LSF identified 62 features from non-contrast-enhanced CTs and achieved root mean square error (RMSE) and correlation coefficient (R) of 1.018 and 0. 9955, respectively, with LOOCV method. SVR-LSF proposed in this study is an optimization regression method. It can not only select a minimal number of features while minimizing mean squared error, but establish both regression and classification model. Adding clinical features to feature set can increase the effective screening of patients undergoing Y-90 radioembolization.