應用卷積神經網路演算法研究TACE術後CT影像腫瘤碘油密度變化分類模型

The Classification of Variety Lipiodol Dense in Tumor on Computed Tomography after TACE by CNN Model

溫漢忠(Han-Chung Wen) ； 蔡明昇(Ming-Sheng Tsai) ； 葉力仁(Lee-Ren Yeh) ； 黃詠暉(Yung- Hui Huang) ； 陳泰賓(Tai-Been Chen)

《中華放射線技術學雜誌》 45卷3期 (2021/09) Pp. 153-158

摘要

目前電腦斷層攝影（Computed Tomography, CT）是醫學造影儀器中具成像時間短、解析度高、應用性廣的臨床影像診斷工具之一。臨床上肝細胞癌（Hepatocellular Carcinoma, HCC）的診斷與追蹤大部分會使用CT進行影像檢查。然而治療HCC除了外科手術切除外，肝動脈化療栓塞術（Transcatheter Arterial Chemoembolization, TACE）亦是常用的方法。本研究主要目的為應用卷積神經網路演算法針對腫瘤碘油CT影像進行TACE治療次數分類之探討。研究方法為回朔性的研究，收集患有肝腫瘤並有使用TACE治療術的個案，後續回診追蹤肝臟血流三相的電腦斷層影像中無顯影劑的CT影像做為樣本。收案條件為（一）TACE術後和每次追蹤的無顯影CT影像；同一個案例可重複施行TACE療程，每一次的療程都被收錄成個別的樣本。（二）共計收集2007年至2019年間條件符合之影像83位HCC肝癌案例，共789筆TACE術後CT影像。依TACE治療次數區分為第一次、第二次、第三次治療組；同時挑選腫瘤最大徑之二維CT影像做為訓練集。利用卷積神經網絡演算法（Convolutional Neural Network Algorithm, CNN）進行TACE治療CT影像特徵萃取，再利用SVM根據影像特徵建立分類模型。本研究結果顯示整體分類準確性有8成以上；未來期望利用卷積神經網絡演算法自動切割或標記CT影像腫瘤碘油所在位置，提供進一步的影像資訊。

關鍵字

肝動脈化療栓塞術 ； 卷積神經網絡演算法 ； 碘油密度

並列摘要

The computed tomography (CT) is one of the clinical imaging diagnostic tools with short scanning time, high resolution and wide applications. In clinical diagnosis and follow-up of Hepatocellular Carcinoma (HCC), CT is commonly used to study the variety of HCC. Meanwhile, transcatheter arterial chemoembolization (TACE) is also a commonly used to therapy HCC. The main purpose of this study is to use convolutional neural network algorithm (CNN) and support vector machine (SVM) to classify the number of TACE after treatments based on CT image with intensity of lipiodol inside. The retrospective was involved in this study. The conditions for collections are: (1) non-contrast CT images after TACE and each follow-up; TACE treatment can be repeated for the same case, and each treatment course is included as an individual sample, (2) a total of 83 HCC liver cancer cases with eligible images from 2007 to 2019 were collected, and a total of 789 CT images after TACE were collected. According to the numbers of TACE, it is divided into the first, second, and third treatment groups. At the same time, the two-dimensional CT image of the largest tumor diameter is selected as the training set. CNN is used to extracted images features for TACE treatment CT tomography. SVM was applied to classify numbers of TACE based on images features. The results of this study show that the classified accuracy is over 80%. In the future, it is expected to use convolutional neural network algorithms to automatically cut or mark the location of tumor lipiodol in CT images to provide further image information.

並列關鍵字

hepatic artery chemoembolization ； convolutional neural network algorithm ； lipiodol density