Please use this identifier to cite or link to this item:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/77478
Title: | 以影像前處理加強全卷積網路於醫療影像輪廓圈選之應用研究 Fully Convolutional Networks with Image Preprocessing for Medical Image Segmentation |
Authors: | Yu-Chun Huang 黃昱鈞 |
Advisor: | 陳正剛(Argon Chen) |
Keyword: | 物件輪廓圈選,醫療影像,機器視覺,卷積型網路, Object Segmentation,Medical Image,Computer Vision,CNN,FCN, |
Publication Year : | 2018 |
Degree: | 碩士 |
Abstract: | 物件輪廓圈選 (Object Segmentation) 為機器視覺中常見的研究與應用。物件輪廓圈選為像素等級的物件識別,要辨別影像中每個像素所屬類別,為機器視覺中難度與精細度較高之應用。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病變 (Lesion) 或感興趣區域 (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。本研究將針對醫學影像的輪廓圈選來研究與討論。
不同資料型態對於影像物件圈選有巨大影響,醫療影像相較於一般影像資料常伴隨高雜訊,導致輪廓難以界定,高斯濾波器是常用來去雜訊之前處理方法之一。醫學影像中的不同組織會在不同影像模式 (Modalities) 下有不同的灰階值表現,如在CT影像中的HU值 (光穿透率) 及超音波影像中的echo值 (音波反射量),直方圖均衡化 (Histogram Equalization) 與直方圖二值化 (Histogram Binarization) 因此常被用來強化一張影像裡不同組織組成的對比度。因此,傳統規則式(Rule-based)輪廓圈選方法透過影像前處理 (Pre-processing),對影像做直方圖分析及不同濾波器 (Filtering) 處理,再將影像像素強度差異放大、取得像素間各方向梯度資訊,最後利用邊緣偵測 (Edge Detection)方法來找尋輪廓。 近年來在電腦運算能力不斷進步下,類神經網路的深度學習如卷積型神經網路 (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,卷積型神經網路中的卷積層濾波器(Convolution filters)必需先隨機初始化,並透過反向傳遞法(Back propagation)更新,但可能因初始化不好、資料量不足、收斂太慢等等原因無法快速學習出我們已知對於問題有幫助的影像特徵。因此本研究將融合醫學影像特性與傳統影像前處理方法,以提升全卷積網路在醫療影像輪廓圈選應用的效率。 本研究首先以直方圖均衡化之觀點出發,透過影像直方圖轉換,將影像直方圖轉成二維頻率圖,進而給予網路學習更直接之影像像素強度分布資訊,接著從影像雜訊過濾及色差強化的觀點出發,將高斯濾波器與梯度資訊的影像前處理結果與原圖一併輸入網路進行學習,藉此觀察影像前處理手法是否可提升醫學影像全卷積網路輪廓圈選之效率與收斂穩定性。高斯濾波器與梯度資訊的前處理計算也引發了這些方法的參數選擇問題,因此本研究同時將高斯濾波與梯度計算設計為可訓練之卷積層濾波器,透過所定義之損失函數更新高斯濾波器參數與梯度之對比強度。 為了驗證上述方法,本研究將以甲狀腺超音波影像與腹部CT影像進行案例分析,分別利用甲狀腺腫瘤超音波輪廓圈選(共1118例)與腹部CT影像骨骼肌輪廓圈選(共215例),探討加入影像直方圖資訊、高斯濾波器、梯度偵測前處理、可訓練式高斯濾波器及可訓練梯度偵測器是否可提升兩類影像輪廓圈選效率及收斂穩定性,進而提出結合影像前處理方法之最佳全卷積網路設計。 Object segmentation is an important research subject in application to computer vision. It can be regard as pixel-level object recognition. The category of each pixel in an image is identified by the algorithm. The pixel-level algorithm is considered more difficult than object-level recognition algorithm. 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) of the lesion for computerized analysis. The contours of the lesion are then defined before subsequent computerized detection or diagnosis. This study will focus on the research and discussion of object segmentation for medical images. Different types of image may require different object segmentation algorithms. Medical images are often accompanied by high noise, which makes the contour definition more difficult. Gaussian filters are one of the commonly used methods to diminish the noise. Different tissues in medical images have different grayscale values under different image modalities, such as light transmittance (HU values ) in CT images and sound wave reflections (echo values ) in ultrasound images. Histogram equalization and histogram binarization are often used to enhance the contrast of different tissues in an image. Conventional rule-based object segmentation methods also utilize various filtering processes on the image to enhance the boundaries between two different tissue parts. Gradient information obtained by filtering in all directions is then used by the edge detection method to find the contour. In recent years, with the continuous advancement of computer computing power, deep learning techniques, such as Convolution Neural Networks (CNN), become readily applicable in many applications. Fully Convolutional Networks (FCN) is one of CNN architectures having good performance on object segmentation. However, Convolution filters in CNN or FCN must be randomly initialized and updated by Back Propagation and consume the updating cycles. There is no guarantee that FCN will learn to obtain those filters most efficient for the purpose of object segmentation. This study combines medical imaging features and traditional image pre-processing methods to improve the efficiency of FCN in learning the medical image object segmentation. This study starts with histogram analysis transforming the image histogram into a two-dimensional frequency map. By this way, FCN will obtain pixel intensity distribution information directly from a two-dimensional frequency image. Then, based on the ideas of image denoising and enhancement, pre-processing results of the Gaussian filter and the gradient information were fed into FCN to test whether the image pre-processing methods can enhance the efficiency of FCN object segmentation. We also attempt to design the FCN such that the parameters of Gaussian filter and gradient operators can be accommodated into the network backpropagation mechanism to become trainable parameters of the networks. In order to verify the above proposed methods, this study will use thyroid ultrasound images and abdominal CT images for case studies. 1118 cases of thyroid nodule segmentation and 215 cases of abdominal CT skeletal muscle segmentation are used for validation. From the study results, it is shown that adding the image histogram information, trainable or non-trainable Gaussian filters and gradient operators can improve the efficiency and convergence stability of the two types of image segmentation problems. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/77478 |
DOI: | 10.6342/NTU201803913 |
Fulltext Rights: | 未授權 |
Appears in Collections: | 工業工程學研究所 |
Files in This Item:
File | Size | Format | |
---|---|---|---|
ntu-107-R05546024-1.pdf Restricted Access | 8.76 MB | Adobe PDF |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.