在U-Net架構上運用遷移式學習於醫療影像之肝組織及腫瘤組織區塊辨識

Abstract

在醫療的疾病診斷上，醫生往往需要透過不同的檢測方式得到資訊，如超音波掃描、抽血、斷層掃描等等。然而在醫療影像的判讀，只能透過肉眼判斷病變的區域，而這需要相當多的經驗累積與專業知識，且容易有誤判的情形產生。隨著影像辨識技術的進步，可透過訓練卷積式類神經網路，讓電腦自動化判讀醫療影像更加精準，藉此輔助醫生的診斷。目前影像分割的研究大多都是以全卷積網路 (FCN) 針對圖片的個別像素做分類，達到影像分割的目的。本文展示一套方法，以U-Net作為基礎架構並且匯入ResNet-50預訓練模型的參數，偵測醫療影像中的肝臟與腫瘤區域。由於資料量的限制，我們也透過調整影像光線對比度的方式、縮放圖片大小與一些基本幾何轉換達到資料擴增的效果。此外我們將訓練兩個模型，其中一個模型首先判別肝臟區域，另一個則負責判斷腫瘤區域。利用第一個模型結果作為判斷腫瘤的輸入資料，以降低其他非肝臟區域造成的雜訊。本文的訓練資料集是來自於MICCAI 2017 LiTS-challenge上的131組CT影像，並有該競賽提供的70組CT影像作為測試資料集。本篇論文所提出的方法可以達到0.71的正確率，和其他2D模型的相關研究相比，在判斷速度與正確性上皆優於其他方法。

For the medical diagnosis of disease, doctors often need to get information through different detection methods, such as medical ultrasound, blood test, and computed tomography. However, for the interpretation of medical images, the region of the lesion can only be distinguished by the naked eyes. This traditional method needs a lot of experiences with professional knowledge, and the possibility of misjudgment exists. With the advancement of image recognition technology, we can make computers automatically interpret medical images more precisely by training the convolutional neural network, and then help judgment by the doctors. Most of the current research on image segmentation is that using the fully convolutional network (FCN) to do pixel-wise classification for image segmentation. This paper presents a method that using U-Net as a basic architecture combining with the weights of ResNet-50 and detecting the areas of liver and tumor in the medical images. Due to the little amount of data, we implement data augmentation methods, including contrast adjustment, rescaling, geometric transformation. In addition, we train two models, one identifies the regions of the liver first, and the other is responsible for regions of the tumor. We use the results from the first model as the input data of the second model to reduce noise caused by other non-liver areas. The training data set of the paper is 131 sets of computed tomography (CT) volumes from MICCAI 2017 Liver and Liver Tumor Segmentation (LiTS) challenge, and there are 70 sets of CT volumes provided by the competition as testing dataset. Our proposed method can reach the accuracy of 0.71, and compared with other 2D models, we have better efficiency and accuracy.