以連續影像全卷積深度學習網路圈選3D醫學影像之研究

Abstract

在機器視覺的研究中，對於圖像的分類問題是常見的實際應用，而其中物件輪廓圈選 (Object Segmentation)為像素等級的物件識別，要辨別影像中每個像素所屬類別，進而產生與原圖相同大小的影像分類結果。而對於醫學影像的電腦輔助偵測及診斷 (Computer-Aided Detection and Diagnosis, CAD)，首要步驟就是必須自動偵測出感興趣區域 (Region of Interest, ROI) 的正確位置並圈選出其輪廓，輪廓定義後才能進行後續的電腦診斷或推論，然而在現今醫療器材的使用中，如電腦斷層攝影（Computed Tomography, CT），利用結合X光與電腦科技的診斷工具將資料組合成身體橫切面的影像，這些橫切面的影像可再進一步重組成精細的3D立體影像，本研究將針對3D醫學立體影像的輪廓圈選來研究與討論。 在現今深度學習的發展下，類神經網路的深度學習架構如卷積型神經網路 (Convolution Neural Networks, CNN) 研究蓬勃發展，其中全卷積網路 (Fully Convolutional Networks, FCN) 更可應用於物件輪廓圈選，然而在上述的模型中，都是在處理單張影像的輪廓圈選，並無法考量三維影像中的連續影像資訊，所以後來其他研究提出了將三維立體影像直接進行物件輪廓圈選的神經網路，但會需要比起單張影像圈選時來的需要非常大量的訓練模型時間及硬體設備。因此，為了避免在訓練三維影像資訊的各種成本但同時需要保留連續影像中的資訊，在本研究中將3D立體影像視為有連續性的單張影像，並以連續數張影像作為訓練資料，進而建構出保留連續性的輪廓圈選模型。 此外，對於輪廓圈選神經網路模型，通常會將下採樣(Down-sampling)過程中的的特徵(feature map)作為上採樣(Up-sampling)額外訓練資訊，然而對於影像訓練時是否必要，在本研究中也將嘗試減少這些資訊來觀察結果。 對於類神經深度學習網路的，除了網路的架構以外，訓練時用來更新神經網路權重的損失函數(Loss function)也是對於訓練很重要的影響因素，由於醫療影像在做輪廓圈選時，時常遇到欲圈選的目標欲背景面積比有很大差異的問題，導致預測像素類別時只要全部都預測背景就能使損失函數值降到一定程度，所以為了讓損失函數更重視的是敏感度(Sensitivity)與特異性(Specificity)的權衡，我們以Dice Loss function作為主要的損失函數，然而Dice Loss主要是應用於單張影像間的結果的評估，因此在本研究中，利用連續影像各自的預測結果，給予不同的權重在損失函數上，進而利用連續性來探討影像的圈選結果。 為了驗證上述方法，本研究與台大醫院合作，將以腹部CT影像進行案例分析，利用腹部CT影像全肝及右肝輪廓圈選(共65例)，探討連續影像模型是否可提升影像輪廓圈選的連續性與圈選成果，進而提出最佳連續影像全卷積網路設計。

In the research of computer vision, the classification for image is a common application. Object segmentation is regard as pixel-level object recognition. It need to identify the category of each pixel in an image and create a classification image which is the same size as original image. For example, the first step in Computer-Aided Detection and Diagnosis (CAD) is to automatically identify the correct position and the region of interest (ROI) for computerized analysis. The contours of the region are then defined before subsequent computerized detection or diagnosis. Computed Tomography (CT), also known as CT-Scan, is a radiological examination that captures detailed images of the human body in three dimensions. It utilizes x-ray photons and digital reconstruction to construct slice of body image, then reconstruct a high resolution 3D images. This study will focus on the research and discussion of object segmentation for 3D medical images. In recent years, with the continuous advancement of deep learning techniques, such as Convolution Neural Networks (CNN), become readily applicable in many applications. Fully Convolutional Networks is known as a deep learning network for object segmentation but usually using for single image. Hence, some research start to construct a neural network for 3D volume segmentation but needs a lot of hard-ware and time to training the network. In order to prevent the training procedure as costly as 3D network but still need to preserve the information of 3D image. We consider the 3D volume as a series of consecutive image. Using different numbers of image as training data to construct the model which preserve the consecutive information. Besides the network architecture, the loss for updating the weight of model is also an import effects. The background category of medical images are often too large than the region we want to segmentation, causing the loss function result will be great if the prediction are all background. In order to put focus on the balance of the sensitivity and the specificity, we using Dice Loss function as our basic loss function. However, this loss function is using for evaluating the result of single image. This study will using different weight of loss for consecutive image and discussing the results by the index of continuity. In order to verify the above proposed methods, this study will use abdominal CT images for case studies. 65 cases abdominal CT whole, left and right liver segmentation are used for validation. From the study results, it is shown that adding the consecutive image result can improve the efficiency and convergence stability of image segmentation problems.