A Study of Cell Segmentation and Classification Techniques of Antinuclear Antibody Images

• Main Authors: Tian-Shing Wang, 王添興
• Other Authors: Tung-Shou Chen
• Format: Others
• Language: zh-TW
• Published: 2008
• Online Access: http://ndltd.ncl.edu.tw/handle/83703216325196991273

碩士 === 國立臺中技術學院 === 資訊科技與應用研究所 === 96 === Autoimmune diseases are caused by antibodies in the human immune system attacking normal cells; these antibodies are, in turn, named anti-self antibodies. Among the anti-self antibodies, the antinuclear antibodies (ANAs) are most often found in different kinds of autoimmune diseases. For immune disease examination, physicians use the location information of ANAs in the cells from ANA test and patients’ symptoms to diagnose the autoimmune disease of patients. The ANA cells could be classified into four patterns based on their locations, namely, diffused, nucleolar, peripheral, and speckled. Currently, indirect immunofluorescent is a popular method used in ANA tests to observe cell patterns under fluorescence microscopy to classify cells. However, the locations of ANAs under fluorescence microscopy are not necessarily clear and aslo require tedious manual visual identifications by several physicians to avoid misidentifications. In recent years, to decrease the error rate of cell classifications, researchers have researched to integrate computer algorithms to assist physicians in automatic cell classifications. In 2002, Perner et al. proposed the two cell classification methods to classify the cell patterns. One is to build a decision tree based on image features, and the other is to build the tree based on expert knowledge. From their experimental results, the two methods present an accuracy rate of 75% and 83.4%, respectively. In 2003, Sack et al. designed a computer-assisted classification scheme which combined Perner et al.’s first method with expert diagnoses. Sack et al.’s proposed method exhibits an accuracy rate of 83%. However, the above image analysis methods require 8p image feature where p was a parameter used in Perner et al.’s first method. When p is 12, the accurate rate of cell classification is the highest; however, computing time is long for processing 96 image features. In this proposal thesis, a cell segmentation algorithm was used to segment cells. Two cell classification methods combined with CLC and LibSVM respectively were built for automated cell classification. In cell segmentation, the cell segmentation algorithm segmented each cell as single cell-blocks; overlapped cells were split into each self-contained cell-block. The method involved the erosion and dilation operations of mathematic morphology for image denoising within separation process. The cell classification method classifies four different cell patterns. Linde-Buzo-Gray algorithm is used where four codebooks will be trained for each pattern. Based on these four codebooks, a new image feature is designed for cells. The CLC and LibSVM were employed for analyzing the image feature of cells and identifying cell patterns. According to experimental results, the proposal methods had accuracy rates of 87.5% and 89.5% on cell classification for both CLC and LibSVM. The proposed methods can improve the accuracy rate of cell identification, and offers a lower-complexity feature calculation method that can save computing time. The proposed methods should have a lower the error rate of misidentifications as opposed to manual identifications, and save time for ANA tests.