Abstract Toward A Completely Automatic Neural Network Based

Toward A Completely Automatic Neural Network Based

Human Chromosome Analysis

Boaz Lerner

University of Cambridge Computer Laboratory

Cambridge CB2 3QG, UK

(email: boaz.lerner@https://www.doczj.com/doc/f018002857.html,)

Abstract

The application of neural networks (NNs) to automatic analysis of chromosome images is investigated in this paper. All aspects of the analysis, namely segmentation, feature description, selection and extraction and finally classification, are studied. As part of the segmentation process, the separation of clusters of partially occluded chromosomes, which is the critical stage that state-of-the-art chromosome analyzers usually fail to accomplish, is performed. First, a moment representation of the image pixels is clustered to create a binary image without a need for threshold selection. Based on the binary image, lines connecting cut points imply possible separations. These hypotheses are verified by a multilayer perceptron (MLP) NN that classifies the two segments created by each separating line. Use of a classification-driven segmentation process gives very promising results without a need for shape modelling or an excessive use of heuristics. In addition, an NN implementation of Sammon’s mapping using principal component based initialization is applied to feature extraction, significantly reducing the dimensionality of the feature space and allowing high classification capability. Finally, by applying MLP based hierarchical classification strategies to a well-explored chromosome database, we achieve a classification performance of 83.6%. This is higher than ever published on this database and an improvement of more than 10% in the error rate. Therefore, basing a chromosome analysis on the NN based techniques that are developed in this research leads toward a completely automatic human chromosome analysis.

Index Terms- Chromosomes, feature extraction, image classification, neural networks, partial occlusion, segmentation

Published in IEEE Transactions on Systems, Man, and Cybernetics Special issue on Artificail Neural Networks, vol. 28, pt. B, pp. 544-552, 1998.

1. Introduction

Human chromosome analysis is an essential task in cytogenetics, especially in prenatal screening and genetic syndrome diagnosis, cancer pathology research and environmentally induced mutagen dosimetry [2]. Cells used for chromosome analysis are taken mostly from amniotic fluid or blood samples. The stage at which the chromosomes are most suitable for analysis is the metaphase (Fig. 1). One of the aims of chromosome analysis is the creation of a karyotype, which is a layout of chromosome images organised by decreasing size in pairs (Fig. 2). The karyotype is obtained by cutting chromosome images from a photograph of a cell, taken using a microscope, and arranging the chromosomes into their appropriate places on the layout according to their visual classification by the cytotechnician. Karyotyping is a useful tool for detecting deviations from normal cell structure. Abnormal cells can have an excess or deficit of a chromosome and/or structural defects, like breaks, fragments or translocations (exchange of genetic material between chromosomes). However, even today, chromosome analysis and karyotyping are performed manually in most cytogenetic laboratories in a repetitive, time consuming and therefore expensive procedure.

Great efforts to develop automatic chromosome classification techniques have been made during the last 25 years. However, all have had limited success and have yielded poor classification results compared with those of a trained cytotechnician [2], [10]. Some of the reasons for the relatively poor performances are the inadequate use of expert knowledge and experience, and insufficient ability to make comparisons and/or elimination among chromosomes within the same metaphase. In addition, the systems always require operator interaction to separate touching and/or overlapping chromosomes and to verify the classification results.

During recent years we have developed [14]-[17] a capability to deal with the most important aspects of chromosome analysis, mainly through the application of NNs. The present study is an additional step to extend this capability by solving the very complex problem of separating images

of touching chromosomes and contributing to a compact representation and a highly accurate chromosome classification. Furthermore, we present here an NN based methodology for a complete automation of all the stages of human chromosome analysis. Therefore, the research aims to (a) resolve, reliably and simply, partial occlusion in a chromosome image, (b) improve chromosome classification performance compared with the current methods, (c) study the benefits of applying NNs to the analysis and (d) present a coherent methodology for automatic chromosome image analysis.

Section 2 of the paper introduces the traditional procedure for computerised chromosome analysis. Section 3 describes our methodology to automatic chromosome analysis while Section 4 presents the results of the application of this methodology.

2. Conventional Computerised Chromosome Analysis

Computerised chromosome analysis consists primarily of pre-processing, segmentation, intermediate processing, feature measurement and selection and finally, classification. The pre-processing stage aims to improve the quality of the cell image by techniques of noise removal, edge enhancement and/or contrast improvement. The segmentation’s purpose is to isolate the metaphase chromosomes from the background, undivided cell nuclei and irrelevant biological materials within the image and from each other. After segmentation, some intermediate processing, as the medial axis transform (MAT) and centromere finding, is usually needed to measure features. The result of feature selection, if applied, is a more condensed representation that still retains most of the important chromosome information. Finally, classification, usually using statistical methods, is performed. This section describes the traditional state-of-the-art techniques to accomplish chromosome analysis whereas Section 3 summarizes our suggested methodology. Although segmentation usually precedes all other stages of image analysis, in our

methodology it is based on them, therefore, we prefer to describe it at the end of the two sections.

2.1 Chromosome description

The use of features rather than the image itself makes the classification procedure easier, faster and more accurate. Different features have been used to describe chromosomes, e.g., features of size (length, area, etc.) [2], [15], [21], band pattern descriptors (e.g., Fourier or Gaussian decomposition of an intensity profile along the chromosome [8]) or global features based on the histogram of gray levels or the 2D Fourier transform components [16]. Among the most discriminative chromosomal features we can find two geometrical features, the length [2], [15], [22] and the centromeric index (the ratio of the short arm length to the whole chromosome length) [2], [10], [15], [22]. Other most discriminative features are the density profile features, which are integral or average intensities along sections perpendicular to the medial axis of the chromosome [2], [10], [15], [22]. The MAT is generally employed to extract these later geometrical and intensity features.

2.2 Feature selection and extraction

It is convenient to distinguish between feature selection in the measurement space and feature selection in a projected space. Techniques belonging to the first category aim to accomplish dimensionality reduction by reducing the number of required measurements and are referred to as feature selection. Techniques belonging to the second category use all the information in the feature vector to yield a lower-dimensional vector and are referred to as feature extraction [4].

Feature selection for classification can be regarded as a search, among all possible transformations, for the one that preserves class separability in the lowest possible dimensional space. However, since the Bayes classifier compares a posteriori probabilities, which are hard to

obtain and their estimates normally have severe biases and variances [6], and since the search for the optimal subset is a combinatorial problem, a few sub-optimal selection techniques have been suggested [4]. Most of these techniques are based on the minimization of criteria of scatter matrices, which are conceptually simple and give systematic feature selection algorithms. These criteria are based on the within-class (W )

W x m x m i i i c i T

=??∈=∑∑ P x i x ()()1(1)

and between-class (B )

B P m m m m i i c i i T

=??=∑1()()(2)

scatter matrices, where m i is the mean of vectors x i in the i th class, m is the mean vector of the mixture distribution, c is the number of classes and P i is the i th class a priori probability.Intuitively, the criterion (e.g., J =trace(W -1.B )) should be larger when the between-class scatter is larger (the classes are distant from each other) and/or when the within-class scatter is smaller (the patterns in each class are concentrated around their centroid). Although many feature selection methods, which are based on different selection techniques and scatter criteria, are known from the literature [4], previous research in chromosomal feature selection has relied only on a very few approaches [9], [15].

In feature extraction, the original features (measurements) are mapped into fewer features which retain the main information of the data structure. A mapping f transforms a pattern y of a d -dimensional measurement space to a pattern x of an m -dimensional space, x =f (y ) and m

{}{})( min )(y g J g y f J =(3)

is determined among all the transformations g [4].

To our knowledge, feature extraction techniques were applied to chromosome analysis only once [17].

2.3 Chromosome classification

Chromosome classification is the most widely investigated stage of chromosome analysis. Over the years several classifiers have been tested, among them: distance and statistical [2], [10], [21], [22], nearest neighbour [10], NN based [7], [14], [27] and fuzzy [29] classifiers. Most of these classifiers have two flaws [10], [16], [22]: 1) lower level of performance than the human expert (~70-80% compared with ~99.7%) and 2) a requirement for an operator interaction to classify misclassified chromosomes. Some of the reasons for these flaws are: 1) inadequate use of the expert knowledge and methodology (e.g., chromosome comparison, elimination), 2) the use of features of limited quality and/or number compared with the very powerful feature synthesis mechanism of the expert brain, 3) the use of the same specific features to represent all chromosomes, and 4) the naive attempt to represent a very complex problem, involving highly context-sensitive patterns, which exhibit natural diversities, by a constrained mathematical and/or engineering framework.

2.4 Segmentation of a chromosome image

Most conventional image segmentation methods are based on either threshold selection, adaptive thresholding, edge detection or matching with a set of prototype shapes [20]. However, almost all of these methods tend to fail or lose accuracy when considering complicated images or those of partially occluded objects [13] as in the case of a chromosome image [16], [18]. Indeed, only a very few studies [1], [16], [18] deal successfully with automatic image segmentation of touching or overlapping chromosomes. Consequently, it is not surprising that in most of the published works concerning chromosome analysis, manually segmented databases are used.

Neither is it surprising to find that almost all the commercial “automatic” chromosome analysis systems are in fact “semi-automatic”and require a continuous interaction of the cytotechnician. In a comprehensive research to separate touching chromosomes, Liang [18] bases his approach on an analysis of concavities along the chromosome shape and on performing a heuristic search for a minimum density path between touching chromosomes. Although very efficient and superior to other techniques, Liang’s method uses a large number of empirical thresholds that might suit images taken from one laboratory (or using a specific preparation technique) but not necessarily others, involves a time-consuming procedure and depends critically on the origin of the split line. In another successful method [1] to separate touching chromosomes, an evaluation of the fit of the separated segments to some prototype shapes is made. However, this method strongly depends on heuristics and thresholds and make use of a multistage complex procedure. Therefore, a complete and efficient automatic segmentation of a chromosome image is a very ambitious goal.

3. A Methodology for Automatic Chromosome Analysis

3.1 Chromosome data

The suggested methodology is applied to two chromosome databases. In the first database, there are images of amniotic fluid cells which were acquired from the Institute of Genetics of Soroka Medical Center, Beer-Sheva, Israel. The images were digitized to yield 512 X 768 pixel pictures where each pixel was represented by 1 byte (256 gray levels). No pre-processing techniques were applied to the images. These images were used both for the automatic segmentation procedure (Section 3.6) and to segment manually five types of chromosomes (the “Soroka5 database”) [15] to evaluate the chromosome features (Section 3.2) and feature selection and extraction paradigms (Sections 3.3 and 3.4). The second database was kindly provided by Dr. Piper of the Medical Research Council in Edinburgh, Scotland. This database

[22] (the “Edinburgh database”) has two main advantages: first, it is huge and hence represents the variety of chromosome images very well. Moreover, its large size is a very crucial advantage when chromosomes are represented by a large number of features. Second, this database has been widely explored by researchers in the field [2], [7], [10], [21], [22], [27], [28] and is therefore very attractive for a comparative study of classification performances (Section 4.2). It contains around 5,500 chromosomes of 125 cells and is based on routinely acquired blood cell images. The characteristics of the two databases are summarized in Table 1.

3.2 Chromosome representation

Most of the chromosomal features described in sub-Section 2.1 were evaluated to select a good chromosome representation [15], [16]. Two geometrical features, the chromosome length (lng.) and centromeric index (ci) and 64 density profile (d.p.) features, which have previously found very discriminative for classification [15], are suggested in this methodology.

3.3 Feature selection

The “knock-out” algorithm [25], which is a sequential backward technique and a simplified scatter criterion which only relies on the within-class scatter matrix, J=trace(W) [15], are employed to find a sub-optimal subset of the 66-dimentional (64 d.p. + ci + lng.) feature space.

3.4 Feature extraction

Different mapping functions and optimization criteria have been suggested in the past for both exploratory data projection and classification. A specific choice of a mapping function and an optimization criterion could be very beneficial to one task but not necessarily to the other. Moreover, the number of features which are required for classification is usually higher than that required for data projection. Therefore, an evaluation of the projection maps and the probability

of correct classification based on several mappings that were applied to the chromosome feature space has been made here. This evaluation includes an NN implementation [19] of Sammon’s mapping [26] (unsupervised nonlinear paradigm) and an MLP [11] (supervised nonlinear), auto-associative NN [11] (unsupervised linear) and principal component (PC) [6] (unsupervised linear) feature extractors.

3.5 Classification

Classification of five types of chromosomes using a monolithic MLP NN was found to be both efficient and superior to that based on the Bayes piecewise linear classifier [14]. However, when classifying all the 24 chromosome types both classifiers yielded comparable results [7]. Therefore, when trying to accomplish a full karyotype using an NN based classifier, other, e.g. hierarchical, strategies may be considered apart from the monolithic MLP. Besides task simplification, which might improve generalization, hierarchical strategies cut training period drastically. When applying hierarchical classification, different classifiers could be trained independently, using independent data sets and/or feature sets, to possibly improve performance and to employ fewer resources. For example, chromosome identification by first classifying the patterns into groups followed by a classification in the groups and into types yields both a desired task decomposition and a compatibility with the common cytogenetic methodology [3], which partitions the twenty-four chromosome types into seven groups. Therefore, a “group classifier” is trained here independently of seven “type classifiers”, each one of them is employed for a different group. Furthermore, training of the “group classifier” could be based on typical “group” features while that of the “type classifiers” on “type” features.

3.6 Automatic segmentation of partially occluded chromosomes

While the problem of segmenting images of isolated chromosomes is easily solved by thresholding, the accuracy of separating images of touching chromosomes is not yet sufficient for a reliable completely automatic procedure [5]. Techniques which are based on threshold selection fail [16], [18], [22] and techniques which are “tailored” to the chromosome shape [1], [18] are complex, multistage procedures which highly depend on heuristics and initial conditions. However, clusters of touching chromosomes exist in almost every chromosome image. Hence, the separation of them is unavoidable to achieve a completely automatic chromosome image analysis, especially in low quality images (e.g., bone marrow) which can have many clusters of touching chromosomes [18].

A new approach called a classification-driven partially occluded object segmentation (CPOOS) method [16], is used here for separating clusters of touching chromosomes. The method aims to achieve a highly accurate separation of touching objects without a need for a tedious, usually unreliable experimentation with threshold selection or the excessive use of heuristics in locating the best separating path between touching objects.

The CPOOS method consists of three main stages. The first stage is K-means clustering(1) of an algebraic moment representation of the image pixels. The algebraic moments, as other statistical features, provide a more comprehensive characterisation than the pixel gray level itself. Although moment invariants are invariant to translation, rotation and scale changes [12], they are more sensitive to noise and sometimes yield inferior classification performance compared with the

algebraic moments [23], therefore the latter are preferred here. The (p,q)-th moment, m

pq

, of a

pixel (x

0,y

) of an image T which is measured in an N x N neighbourhood around the pixel is:

(1) The term “cluster” is usually used to describe both the aggregation of partially occluded objects [18] and the operation of grouping similar patterns (pixels) into groups (clustering) [20].

pq p q y N x N m x y x x y y T x y (,)()()(,)000011=??==∑∑ , p,q =0,1,2, (4)

It has been proved [12] that the double moment sequence, m pq , is uniquely determined by T and vice versa . Only very few moments are required here for two reasons. First, a large number of moments have not been found to be superior to a small number [23]. Second, we look for a rapid implementation, hence the smaller the number of moments that provides fair results the better.Therefore, only two moments (m 01 and m 02) which are calculated in a 3x3 neighbourhood around the central pixel are extracted and found to provide a robust B/W segmentation (K=2).

In the second stage, clusters of touching chromosomes are identified by their size and following a failure to classify them to one of the chromosome types. A cluster is identified if the probability of classifying it to each of the types is less than the inverse of the number of types.The contour of a cluster is found based on the 8-connected neighborhood of each contour pixel and tracked so that the image interior is located to the left of the contour. Following Rosenfeld and Kak [24], we define the left and right “k -slope” of a digital contour S =(p 0, .., p n ), where p i and p i+1 are adjacent points, as the directions from p i to p i-k and from p i to p i+k , respectively,where k ≥1. Based on this definition, the “k -curvature” (the angle between left and right “k -slopes”) of each contour point is computed and points are ordered by their curvatures. The r most concave points are selected and suggested as cut points to draw possible separating lines between touching chromosomes. A point, among these r potential cut points, which is closer than a distance s from a more acutely concave point is eliminated. The values of these experimentally found parameters remain unchanged throughout the experiments reported.

In the final stage, each pair of cut points creates a potential separating line. When the line connecting cut point(i ) and cut point(j ) is hypothesized, the two segments created by drawing this line are suspected of being the touching chromosomes. Each of the segments is rotated based on its first principal axis to be aligned to the x =0 axis, and the “straight” segment is transformed to

its medial axis (MA) [15] to extract the chromosomal features (ci, lng., and the 64-dimensional d.p. features of Section 3.2). Following the normalization of the feature vectors of the training segments (see later) to zero mean and unit variance the density profile feature vectors (only) are projected onto their principal axes. This feature extraction stage enables the use of approximately 90% of the variance in the density profiles with only one sixteenth of the number of original features (see later in Fig. 3) and without any performance degradation. An MLP, trained by the backpropogation (BP) learning algorithm [11], classifies the two reduced vectors (four eigenfeatures plus two geometrical features) and thereby achieves hypothesis verification and the selection of the “correct” separation line. The MLP and the BP provide simple but very reliable means of classification. Additionally, when implemented in a modest configuration which is therefore trained for only very short periods they accomplish a very cost-effective “tool” for hypothesis verification. The classifier is trained beforehand based on a priori knowledge of the identity of the two types of chromosomes which compose the cluster. This knowledge is gained following the classifiation of the isolated chromosomes in the beginning and the application of a simple elimination criterion. We assume that only one cluster of touching chromosomes exists in the image. When the image consists of several clusters and/or clusters of more than two chromosomes, the method would be applied hierarchically using complementary heuristics.

The classification is based on three classes. The first two are the two expected types of chromosomes which compose the cluster (the “wholes”) and the third is of those images created by any arbitrary “wrong” separating line (the “brokens”). The expected output vectors of the first two classes are (100) and (010) (or in opposite order for the opposite order of inputs) for a “correct” separating line and (001) and (001) for a “wrong” one. During the test, the first output value of the first output vector and the second output value of the second output vector are averaged to yield a score, which is an estimation of the average maximum a posteriori probability of the two input vectors. Therefore, a line that creates segments which are largely (or entirely)

composed of the expected chromosomes yields a high score where a line between two “wrong”cut points yields a low score. The highest score which is assigned by the classifier indicates the “correct” separating line and thereby verifies the corresponding hypothesis.

In addition, a “geometrical constraint” is obtatined from the training set using the maximum and minimum values of the length and centromeric index. A separating line in a test image that creates two segments such that at least one of them yields a length or a centromeric index which is either larger than one of the above maxima or smaller than one of the above minima is rejected and is not checked by the classifier.

4. Experiments and Results

4.1 Feature measurement, selection and extraction

Although global features are very easily computed and suitable for hardware implementation [16], a combination of geometrical ((lng.) and (ci)) and intensity (d.p.) features (Section 3.2) is shown in Table 2 to provide better chromosome discrimination when classifying five types of chromosomes (similar conclusions are made in [22] for 24 types). The latter features are used throughout this study and measuring them is part of the suggested methodology. In all the experiments of this sub-section, a two-layer perceptron trained by the backpropagation (BP) algorithm is employed as a classifier.

When the “knock-out” algorithm and the scatter criterion of Section 3.3 are applied to the “Soroka5 database” to choose a feature subset from the above 66 features, the length and centromeric index of the chromosome and points of the d.p. which correspond to the chromosome upper tip are selected as the most significant features [15].

When different mapping paradigms are applied to the chromosome features, superior paradigms for exploratory data projection are often found to be superior paradigms for classification and vice versa [17]. An experiment to evaluate the four mappings of Section 3.4 is

performed with a data set of 300 patterns of three types of chromosomes (types “13”, “19” and “x”), using 100 patterns of each type. The patterns are derived from the “Soroka5 database”, represented by 64 d.p. features and mapped into 1 to 10 dimensional patterns. The experiment is repeated with ten randomly initialized classifiers and fourty randomly chosen training and test sets derived from the database. The average probability of correct classification of the test set is shown in Fig. 3 (for details on the paradigm parameters see [17]). Only very few features (around one sixteenth of the original number) are required to achieve almost the ultimate performance, therefore contributing to a substantial reduction of the feature space dimensionality. Although Sammon’s mapping aims at data visualization, we can extract an arbitrary number of features by the mapping and therefore apply it also to classification. Fig. 3 reveals an impressive chromosome classification performance based on the unsupervised Sammon’s mapping which is comparable with that based on the supervised MLP feature extractor, and superior to those of the auto-associative NN and PC feature extractors. Furthermore, we initialise the input-hidden weight matrix of the NN implementation of the mapping by the eigenvectors of the covariance matrix of the data [17]. Besides improved classification capability, the eigenvector based initialisation of Sammon’s mapping provides a reduced training period, a lower mapping error and requires fewer experiments compared with the random initialisation.

4.2 Classification

The classification accuracy of two monolithic configurations and two hierarchical strategies, all of them based on an MLP, is compared for the “Edinburgh database” and for classifying the full karyotype (24 chromosome types). The feature space is 17-dimensional and it includes 15 selected out of the 64 d.p. features along with the length and centromeric index of the chromosome. Classification performance is the average of two experiments, where in the first training is made on one half of the data and testing on the other half and in the second the

procedure is reversed (i.e., the cross-validation technique). This technique is common in the cytogenetic community [21], [22], and therefore it is also applied here for a comparison.

In the first hierarchical strategy the chromosomes are first classified into seven groups and then in each of the seven groups (Section 3.5). The same features are used by all the classifiers. The configurations of the classifiers are: 17:hidden:7 for the “group classifier” and 17:hidden:output for the “type classifiers”, where “hidden” is the number of hidden units (selected to prevent overtraining) and “output” is 2 to 8 (according to the number of chromosome types in each group). Table 3 shows the probabilities of correct classification of the test set for each of the eight classifiers (one “group classifier” and seven (A to G) “type classifiers”). The first row of Table 3 shows the population of chromosomes in the seven groups. The second and third rows give, respectively, the above probabilities of the “group classifier” and “type classifiers”. The classification probabilities of the “type classifiers” are calculated for only those chromosomes that are correctly classified by the “group classifier”. The overall performance of the first hierarchical classifier is calculated, based on the relative population of chromosomes in the groups, to be 83.6%. When this experiment is repeated using only the two geometrical (ci + lng.) features to classify chromosomes into their groups and only the intensity (d.p.) features to classify chromosomes into their types, a performance degradation of all the eight classifiers (~2-10%) is observed.

In the second hierarchical strategy, a “group classifier” is trained using only the geometrical features (configuration of 2:hidden:7), and its outputs, as well as the fifteen selected d.p. features, are used as the inputs of a “type classifier” (configuration of 22:hidden:24). The probability of correct classification to “groups” is 89.4% and that to “types”, which is also the overall probability, is 81.6%, which is lower by 1.3% than that reported by Graham et al. [7] for a similar strategy.

Table 4 compares our classification performances using the monolithic classifiers (two and three layer perceptron in columns 1 and 2, respectively) and the two hierarchical strategies (columns 3 and 4) with the published best classification performances [2], [7], [22], [28]. Our classification methods, as the maximum likelihood classifier (column 5), belong to the context-free classification methods, in which the data set is classified as is and without a posteriori rearrangement of the chromosomes. Context-dependent classification procedures (columns 6 and 7), on the other hand, make use of a priori knowledge that each normal cell has exactly two chromosomes from each type. A classification procedure which results in a different arrangement of the chromosomes (e.g., three or more chromosomes of a specific cell are allocated to the same type) is followed by rearranging the chromosomes to meet the expected pattern [22], [28]. These classification procedures are found to provide a higher probability of correct classification by about 3% than context-free classification [2], [22], [28] but they are only useful if the cells are known to be taken from a normal sample. The first context-dependent classification method (column 6) is based on the maximum likelihood classifier (column 5) and a penalty for “implausible” assignments [22]. The second method (column 7), called the “transportation method” [28], finds a globally optimal solution to the constrained allocation problem. A third context-dependent classifier based on a probabilistic NN [27] yields very similar results to the first two. Table 4 reveals that the first hierarchical strategy (column 3) is slightly superior to the other context-free classification methods (columns 1, 2, 4 and 5) and even comparable with the context-dependent classification methods (columns 6 and 7).

4.3 Automatic segmentation

One of the novelties of using the CPOOS method for separating partially occluded chromosomes is that it is a classification-driven segmentation procedure. A necessary condition

for a successful application of the CPOOS method is, therefore, a demonstrated chromosome classification capability, such as the one described so far.

Only sub-images of cell images that include clusters of touching chromosomes are selected and analysed here. Training and test images are gathered for different combinations of two touching chromosomes. Since there are many as p(p-1)/2 such combinations (p=24) and there is no standard database of cell images, we are able to collect only very few images with the same combination of touching chromosomes. Therefore, in most of our experiments only one image is used for training and another one for the test, although there are a few examples where we have two training images. Using each training image and the cluster’s potential separating lines we create segment images of arbitrary cluster separation which define the data set of images of the third class (the “brokens”) for a specific combination. The images of the first two classes (the “wholes”) are collected from a manually segmented database (sub-Section 3.1). The two sets (“brokens” and “wholes”) are combined to one data set of around 300 vectors (~100 from each class) and the procedure repeats itself for all the p(p-1)/2 combinations (networks). The 66-dimensional feature vectors, one from each image of each class, are measured, normalised and projected into the first four principal axes (the quickest mapping among the four of Fig. 3) to define the training set for the classifier. During the test, the classifier is similarly used to verify each hypothesis which is posed by each separating line in the test image. In addition, these lines are checked to satisfy the “geometrical constraint”.

To cover all the possible combinations of two touching chromosomes p(p-1)/2 networks are required. However, each one of these networks is of a very modest configuration (6:2:3) and is therefore trained for a very short period (100 epochs). Moreover, since the classifier training is done beforehand, the verification procedure (test) is very rapid and only involves a multiplication of a 2x6 feature matrix by two weight (6x2 and 2x3) matrices.

Fig. 4 shows two examples of training and test images of clusters of touching chromosomes along with the cluster contours and cut points of the test images. The two training images (Fig. 4a) are c107 (top) and c159 (bottom) and the two test images (Fig. 4b) are c121 (top ) and c163 (bottom). The clusters in the first and second examples are composed of chromosome types “2”and “x” and “2” and “4”, respectively.

Tables 5 and 6 outline the separation results for each test image of Fig. 4 using lines connecting each pair of cut points. These results are represented by the classifier score and the satisfaction (or not) of the “geometrical constraint” (sub-Section 3.6). When a separating line does not satisfy the constraint no use of the classifier is made (not applicable (NA) in Tables 5 and 6). In the first test image (c121, Table 5) three lines begin at the same cut point (#8) and end at a relatively very close cut points (#1, #2, #3). However, the classifier ranks the “correct”separating line (2-8) higher than the other two lines (1-8 and 3-8) and with a score close to 1, since it is the only line which creates two segments that are classified as the chromosomes. Lines 1-8 and 3-8, however, create segments of which only one is classified as a chromosome. All the other lines create two segments which are both classified as belonging to the third class (and hence yield a zero score) or do not satisfy the “geometrical constraint”. In the second case (c163, Table 6), a few lines that are connected to point #5 pass near the second correct point (#3) but create slightly different segments, and therefore fail to outperform the score of the “correct” line (3-5). In this test image fewer lines are rejected by the “geometrical constraint” compared with the first image. When applied to a relatively small database of 46 human cell images, the CPOOS method achieves over 90% probability of correct segmentation when it is allowed to reject 8.7% of the images.

5. Toward a Completely Automatic Neural Network Based Human Chromosome Analysis-

a Discussion

The research has had four major goals: (a) to accomplish a reliable but simple segmentation of a chromosome image that contains partial occlusion, (b) to try to improve chromosome classification performance compared with the current methods, (c) to study the benefits of applying NNs to all stages of chromosome analysis and (d) to present a coherent methodology toward a completely automatic chromosome image analysis.

Segmentation of partially occluded chromosomes has been efficiently accomplished in this research using the CPOOS method. First, clustering of the image pixels provides a separation of chromosomes and chromosome clusters from background without a need to search for an optimum threshold. Thereafter, cut points are located at the most concave points along a cluster curvature. Finally, an MLP is used to score and verify hypotheses, which are lines connecting each pair of cut points, by classifying the two segments created by each line. Using this method, modeling of the object’s shape or curvature is not necessary nor is the traditional excessive use of heuristics and pre-processing. Moreover, being a classification-driven segmentation procedure, the CPOOS method is able to apply the (a priori known) information necessary for object classification during the segmentation and thereby to achieve very promising results. The next step in the research is the modification of the CPOOS method to handle more than one cluster in an image and clusters of more than two chromosomes, through hierarchical application of the method combined with suitable heuristics. A modified system such as this could automatically and very reliably segment and classify most of the normal cells (about 97% of the cells) in each cytogenetic laboratory, leaving the human expert only the most difficult cases and the abnormal cells. Although the CPOOS method is applied here to chromosome analysis we believe it is suitable to images with partial occlusion coming from other applications, as well.

Efficient chromosome description is found to depend on a few geometrical features and on the density profile of the chromosome. Global features, although very easily computed, fail to achieve the high discriminative power of the former features. Furthermore, several NN based feature extraction techniques are found, in this research, to be very useful in reducing the feature space dimensionality while retaining high chromosome classification performance.

MLP based hierarchical classification strategies reduce the training period and improve human chromosome classification beyond any known method. Using a very well explored database, an NN based hierarchical classifier has classified 5,500 chromosomes with a success rate of 83.6% compared with 81.7% of a maximum likelihood classifier [2], [21], [22] and 82.9% of another NN based classifier [7]. This is an improvement of more than ten and 4% in the error rate compared with the statistical and the other NN classifier, respectively.

In summary, NNs have been applied to perform all major stages of automatic human chromosome analysis, namely feature extraction, image segmentation and classification. In addition, the integration of these stages into one coherent system guarantees the accomplishment of the research goals and leads toward a completely automatic human chromosome analysis_a crucial necessity in numerous clinical fields, such as prenatal diagnosis, cancer research and detection of radiation-induced aberrations.

Finally, although applied to human chromosome analysis, our methodology is well suited to other real-world image analysis applications and therefore, contributes to the generic research of image analysis.

Acknowledgement The author would like to thank Dr. Romem, previously of the Soroka Medical Center, Beer-Sheva, Israel, and Dr. Piper of the Medical Research Council, Edinburgh, UK, for providing the “Soroka” and “Edinburgh” databases, respectively. This work was supported in part by the Paul Ivanier Center for Robotics and Production Management, Ben-Gurion University, Beer-Sheva, Israel.

公文写作规范格式

商务公文写作目录 一、商务公文的基本知识 二、应把握的事项与原则 三、常用商务公文写作要点 四、常见错误与问题

一、商务公文的基本知识 1、商务公文的概念与意义 商务公文是商业事务中的公务文书，是企业在生产经营管理活动中产生的，按照严格的、既定的生效程序和规范的格式而制定的具有传递信息和记录作用的载体。规范严谨的商务文书，不仅是贯彻企业执行力的重要保障，而且已经成为现代企业管理的基础中不可或缺的内容。商务公文的水平也是反映企业形象的一个窗口，商务公文的写作能力常成为评价员工职业素质的重要尺度之一。 2、商务公文分类：（1）根据形成和作用的商务活动领域，可分为通用公文和专用公文两类（2）根据内容涉及秘密的程度，可分为对外公开、限国内公开、内部使用、秘密、机密、绝密六类（3）根据行文方向，可分为上行文、下行文、平行文三类（4）根据内容的性质，可分为规范性、指导性、公布性、陈述呈请性、商洽性、证明性公文（5）根据处理时限的要求，可分为平件、急件、特急件三类（6）根据来源，在一个部门内部可分为收文、发文两类。 3、常用商务公文： （1）公务信息：包括通知、通报、通告、会议纪要、会议记录等 （2）上下沟通：包括请示、报告、公函、批复、意见等 （3）建规立矩：包括企业各类管理规章制度、决定、命令、任命等； （4）包容大事小情：包括简报、调查报告、计划、总结、述职报告等； （5）对外宣传：礼仪类应用文、领导演讲稿、邀请函等； （6）财经类：经济合同、委托授权书等； （7）其他：电子邮件、便条、单据类（借条、欠条、领条、收条）等。 考虑到在座的主要岗位，本次讲座涉及请示、报告、函、计划、总结、规章制度的写作，重点谈述职报告的写作。 4、商务公文的特点： （1）制作者是商务组织。（2）具有特定效力，用于处理商务。 （3）具有规范的结构和格式，而不像私人文件靠“约定俗成”的格式。商务公文区别于其它文章的主要特点是具有法定效力与规范格式的文件。 5、商务公文的四个构成要素： （1）意图：主观上要达到的目标 (2）结构：有效划分层次和段落，巧设过渡和照应 (3）材料：组织材料要注意多、细、精、严 (4) 正确使用专业术语、熟语、流行语等词语，适当运用模糊语言、模态词语与古词语。 6、基本文体与结构 商务文体区别于其他文体的特殊属性主要有直接应用性、全面真实性、结构格式的规范性。其特征表现为：被强制性规定采用白话文形式，兼用议论、说明、叙述三种基本表达方法。商务公文的基本组成部分有：标题、正文、作者、日期、印章或签署、主题词。其它组成部分有文头、发文字号、签发人、保密等级、紧急程度、主送机关、附件及其标记、抄送机关、注释、印发说明等。印章或签署均为证实公文作者合法性、真实性及公文效力的标志。 7、稿本 （1）草稿。常有“讨论稿”“征求意见稿”“送审稿”“草稿”“初稿”“二稿”“三稿”等标记。（2）定稿。是制作公文正本的标准依据。有法定的生效标志（签发等）。（3）正本。格式正规并有印章或签署等表明真实性、权威性、有效性。（4）试行本。在试验期间具有正式公文的法定效力。（5）暂行本。在规定

关于会议纪要的规范格式和写作要求

关于会议纪要的规范格式和写作要求 一、会议纪要的概念 会议纪要是一种记载和传达会议基本情况或主要精神、议定事项等内容的规定性公文。是在会议记录的基础上，对会议的主要内容及议定的事项，经过摘要整理的、需要贯彻执行或公布于报刊的具有纪实性和指导性的文件。 会议纪要根据适用范围、内容和作用，分为三种类型： 1、办公会议纪要(也指日常行政工作类会议纪要)，主要用于单位开会讨论研究问题，商定决议事项，安排布置工作，为开展工作提供指导和依据。如，xx学校工作会议纪要、部长办公会议纪要、市委常委会议纪要。 2、专项会议纪要(也指协商交流性会议纪要)，主要用于各类交流会、研讨会、座谈会等会议纪要，目的是听取情况、传递信息、研讨问题、启发工作等。如，xx县脱贫致富工作座谈会议纪要。 3、代表会议纪要(也指程序类会议纪要)。它侧重于记录会议议程和通过的决议，以及今后工作的建议。如《××省第一次盲人聋哑人代表会议纪要》、《xx市第x次代表大会会议纪要》。 另外，还有工作汇报、交流会，部门之间的联席会等方面的纪要，但基本上都系日常工作类的会议纪要。 二、会议纪要的格式 会议纪要通常由标题、正文、结尾三部分构成。

1、标题有三种方式：一是会议名称加纪要，如《全国农村工作会议纪要》;二是召开会议的机关加内容加纪要，也可简化为机关加纪要，如《省经贸委关于企业扭亏会议纪要》、《xx组织部部长办公会议纪要》;三是正副标题相结合，如《维护财政制度加强经济管理——在xx部门xx座谈会上的发言纪要》。 会议纪要应在标题的下方标注成文日期，位置居中，并用括号括起。作为文件下发的会议纪要应在版头部分标注文号，行文单位和成文日期在文末落款(加盖印章)。 2、会议纪要正文一般由两部分组成。 (1)开头，主要指会议概况，包括会议时间、地点、名称、主持人，与会人员，基本议程。 (2)主体，主要指会议的精神和议定事项。常务会、办公会、日常工作例会的纪要，一般包括会议内容、议定事项，有的还可概述议定事项的意义。工作会议、专业会议和座谈会的纪要，往往还要写出经验、做法、今后工作的意见、措施和要求。 (3)结尾，主要是对会议的总结、发言评价和主持人的要求或发出的号召、提出的要求等。一般会议纪要不需要写结束语，主体部分写完就结束。 三、会议纪要的写法 根据会议性质、规模、议题等不同，正文部分大致可以有以下几种写法： 1、集中概述法(综合式)。这种写法是把会议的基本情况，讨

titlesec宏包使用手册

titlesec&titletoc中文文档 张海军编译 makeday1984@https://www.doczj.com/doc/f018002857.html, 2009年10月 目录 1简介,1 2titlesec基本功能,2 2.1.格式,2.—2.2.间隔, 3.—2.3.工具,3. 3titlesec用法进阶,3 3.1.标题格式,3.—3.2.标题间距, 4.—3.3.与间隔相关的工具, 5.—3.4.标题 填充,5.—3.5.页面类型,6.—3.6.断行,6. 4titletoc部分,6 4.1.titletoc快速上手,6. 1简介 The titlesec and titletoc宏包是用来改变L A T E X中默认标题和目录样式的，可以提供当前L A T E X中没有的功能。Piet van Oostrum写的fancyhdr宏包、Rowland McDonnell的sectsty宏包以及Peter Wilson的tocloft宏包用法更容易些；如果希望用法简单的朋友，可以考虑使用它们。 要想正确使用titlesec宏包，首先要明白L A T E X中标题的构成，一个完整的标题是由标签+间隔+标题内容构成的。比如： 1.这是一个标题，此标题中 1.就是这个标题的标签，这是一个标签是此标题的内容，它们之间的间距就是间隔了。 1

2titlesec基本功能 改变标题样式最容易的方法就是用几向个命令和一系列选项。如果你感觉用这种方法已经能满足你的需求，就不要读除本节之外的其它章节了1。 2.1格式 格式里用三组选项来控制字体的簇、大小以及对齐方法。没有必要设置每一个选项，因为有些选项已经有默认值了。 rm s f t t md b f up i t s l s c 用来控制字体的族和形状2，默认是bf,详情见表1。 项目意义备注（相当于） rm roman字体\textrm{...} sf sans serif字体\textsf{...} tt typewriter字体\texttt{...} md mdseries(中等粗体)\textmd{...} bf bfseries(粗体)\textbf{...} up直立字体\textup{...} it italic字体\textit{...} sl slanted字体\textsl{...} sc小号大写字母\textsc{...} 表1:字体族、形状选项 bf和md属于控制字体形状，其余均是切换字体族的。 b i g medium s m a l l t i n y（大、中、小、很小） 用来标题字体的大小，默认是big。 1这句话是宏包作者说的，不过我感觉大多情况下，是不能满足需要的，特别是中文排版，英文 可能会好些！ 2L A T E X中的字体有5种属性：编码、族、形状、系列和尺寸。 2

毕业论文写作要求与格式规范

毕业论文写作要求与格式规范 关于《毕业论文写作要求与格式规范》，是我们特意为大家整理的，希望对大家有所帮助。 (一)文体 毕业论文文体类型一般分为：试验论文、专题论文、调查报告、文献综述、个案评述、计算设计等。学生根据自己的实际情况，可以选择适合的文体写作。 (二)文风 符合科研论文写作的基本要求：科学性、创造性、逻辑性、

实用性、可读性、规范性等。写作态度要严肃认真，论证主题应有一定理论或应用价值;立论应科学正确，论据应充实可靠，结构层次应清晰合理，推理论证应逻辑严密。行文应简练，文笔应通顺，文字应朴实，撰写应规范，要求使用科研论文特有的科学语言。 (三)论文结构与排列顺序 毕业论文，一般由封面、独创性声明及版权授权书、摘要、目录、正文、后记、参考文献、附录等部分组成并按前后顺序排列。 1.封面：毕业论文(设计)封面具体要求如下： (1)论文题目应能概括论文的主要内容，切题、简洁，不超过30字，可分两行排列;

(2)层次：大学本科、大学专科 (3)专业名称：机电一体化技术、计算机应用技术、计算机网络技术、数控技术、模具设计与制造、电子信息、电脑艺术设计、会计电算化、商务英语、市场营销、电子商务、生物技术应用、设施农业技术、园林工程技术、中草药栽培技术和畜牧兽医等专业，应按照标准表述填写; (4)日期：毕业论文(设计)完成时间。 2.独创性声明和关于论文使用授权的说明：需要学生本人签字。 3.摘要：论文摘要的字数一般为300字左右。摘要是对论文的内容不加注释和评论的简短陈述，是文章内容的高度概括。主要内容包括：该项研究工作的内容、目的及其重要性;所使用的实验方法;总结研究成果，突出作者的新见解;研究结论及其意义。摘要中不列举例证，不描述研究过程，不做自我评价。

公文格式规范与常见公文写作

公文格式规范与常见公文写作 一、公文概述与公文格式规范 党政机关公文种类的区分、用途的确定及格式规范等，由中共中央办公厅、国务院办公厅于2012年4月16日印发，2012年7月1日施行的《党政机关公文处理工作条例》规定。之前相关条例、办法停止执行。 （一）公文的含义 公文，即公务文书的简称，属应用文。 广义的公文，指党政机关、社会团体、企事业单位，为处理公务按照一定程序而形成的体式完整的文字材料。 狭义的公文，是指在机关、单位之间，以规范体式运行的文字材料，俗称“红头文件”。 ?（二）公文的行文方向和原则 ?、上行文下级机关向上级机关行文。有“请示”、“报告”、和“意见”。 ?、平行文同级机关或不相隶属机关之间行文。主要有“函”、“议案”和“意见”。 ?、下行文上级机关向下级机关行文。主要有“决议”、“决定”、“命令”、“公报”、“公告”、“通告”、“意见”、“通知”、“通报”、“批复”和“会议纪要”等。 ?其中，“意见”、“会议纪要”可上行文、平行文、下行文。?“通报”可下行文和平行文。 ?原则： ?、根据本机关隶属关系和职权范围确定行文关系 ?、一般不得越级行文 ?、同级机关可以联合行文 ?、受双重领导的机关应分清主送机关和抄送机关 ?、党政机关的部门一般不得向下级党政机关行文 ?(三) 公文的种类及用途 ?、决议。适用于会议讨论通过的重大决策事项。 ?、决定。适用于对重要事项作出决策和部署、奖惩有关单位和人员、变更或撤销下级机关不适当的决定事项。

?、命令（令）。适用于公布行政法规和规章、宣布施行重大强制性措施、批准授予和晋升衔级、嘉奖有关单位和人员。 ?、公报。适用于公布重要决定或者重大事项。 ?、公告。适用于向国内外宣布重要事项或者法定事项。 ?、通告。适用于在一定范围内公布应当遵守或者周知的事项。?、意见。适用于对重要问题提出见解和处理办法。 ?、通知。适用于发布、传达要求下级机关执行和有关单位周知或者执行的事项，批转、转发公文。 ?、通报。适用于表彰先进、批评错误、传达重要精神和告知重要情况。 ?、报告。适用于向上级机关汇报工作、反映情况，回复上级机关的询问。 ?、请示。适用于向上级机关请求指示、批准。 ?、批复。适用于答复下级机关请示事项。 ?、议案。适用于各级人民政府按照法律程序向同级人民代表大会或者人民代表大会常务委员会提请审议事项。 ?、函。适用于不相隶属机关之间商洽工作、询问和答复问题、请求批准和答复审批事项。 ?、纪要。适用于记载会议主要情况和议定事项。?（四）、公文的格式规范 ?、眉首的规范 ?（）、份号 ?也称编号，置于公文首页左上角第行，顶格标注。“秘密”以上等级的党政机关公文，应当标注份号。 ?（）、密级和保密期限 ?分“绝密”、“机密”、“秘密”三个等级。标注在份号下方。?（）、紧急程度 ?分为“特急”和“加急”。由公文签发人根据实际需要确定使用与否。标注在密级下方。 ?（）、发文机关标志（或称版头） ?由发文机关全称或规范化简称加“文件”二字组成。套红醒目，位于公文首页正中居上位置（按《党政机关公文格式》标准排

ctex 宏包说明 ctex

ctex宏包说明 https://www.doczj.com/doc/f018002857.html,? 版本号：v1.02c修改日期：2011/03/11 摘要 ctex宏包提供了一个统一的中文L A T E X文档框架，底层支持CCT、CJK和xeCJK 三种中文L A T E X系统。ctex宏包提供了编写中文L A T E X文档常用的一些宏定义和命令。 ctex宏包需要CCT系统或者CJK宏包或者xeCJK宏包的支持。主要文件包括ctexart.cls、ctexrep.cls、ctexbook.cls和ctex.sty、ctexcap.sty。 ctex宏包由https://www.doczj.com/doc/f018002857.html,制作并负责维护。 目录 1简介2 2使用帮助3 2.1使用CJK或xeCJK (3) 2.2使用CCT (3) 2.3选项 (4) 2.3.1只能用于文档类的选项 (4) 2.3.2只能用于文档类和ctexcap.sty的选项 (4) 2.3.3中文编码选项 (4) 2.3.4中文字库选项 (5) 2.3.5CCT引擎选项 (5) 2.3.6排版风格选项 (5) 2.3.7宏包兼容选项 (6) 2.3.8缺省选项 (6) 2.4基本命令 (6) 2.4.1字体设置 (6) 2.4.2字号、字距、字宽和缩进 (7) ?https://www.doczj.com/doc/f018002857.html, 1

1简介2 2.4.3中文数字转换 (7) 2.5高级设置 (8) 2.5.1章节标题设置 (9) 2.5.2部分修改标题格式 (12) 2.5.3附录标题设置 (12) 2.5.4其他标题设置 (13) 2.5.5其他设置 (13) 2.6配置文件 (14) 3版本更新15 4开发人员17 1简介 这个宏包的部分原始代码来自于由王磊编写cjkbook.cls文档类，还有一小部分原始代码来自于吴凌云编写的GB.cap文件。原来的这些工作都是零零碎碎编写的，没有认真、系统的设计，也没有用户文档，非常不利于维护和改进。2003年，吴凌云用doc和docstrip工具重新编写了整个文档，并增加了许多新的功能。2007年，oseen和王越在ctex宏包基础上增加了对UTF-8编码的支持，开发出了ctexutf8宏包。2009年5月，我们在Google Code建立了ctex-kit项目1，对ctex宏包及相关宏包和脚本进行了整合，并加入了对XeT E X的支持。该项目由https://www.doczj.com/doc/f018002857.html,社区的开发者共同维护，新版本号为v0.9。在开发新版本时，考虑到合作开发和调试的方便，我们不再使用doc和docstrip工具，改为直接编写宏包文件。 最初Knuth设计开发T E X的时候没有考虑到支持多国语言，特别是多字节的中日韩语言。这使得T E X以至后来的L A T E X对中文的支持一直不是很好。即使在CJK解决了中文字符处理的问题以后，中文用户使用L A T E X仍然要面对许多困难。最常见的就是中文化的标题。由于中文习惯和西方语言的不同，使得很难直接使用原有的标题结构来表示中文标题。因此需要对标准L A T E X宏包做较大的修改。此外，还有诸如中文字号的对应关系等等。ctex宏包正是尝试着解决这些问题。中间很多地方用到了在https://www.doczj.com/doc/f018002857.html,论坛上的讨论结果，在此对参与讨论的朋友们表示感谢。 ctex宏包由五个主要文件构成：ctexart.cls、ctexrep.cls、ctexbook.cls和ctex.sty、ctexcap.sty。ctex.sty主要是提供整合的中文环境，可以配合大多数文档类使用。而ctexcap.sty则是在ctex.sty的基础上对L A T E X的三个标准文档类的格式进行修改以符合中文习惯，该宏包只能配合这三个标准文档类使用。ctexart.cls、ctexrep.cls、ctexbook.cls则是ctex.sty、ctexcap.sty分别和三个标准文档类结合产生的新文档类，除了包含ctex.sty、ctexcap.sty的所有功能，还加入了一些修改文档类缺省设置的内容（如使用五号字体为缺省字体）。 1https://www.doczj.com/doc/f018002857.html,/p/ctex-kit/

文档书写格式规范要求

学生会文档书写格式规范要求 目前各部门在日常文书编撰中大多按照个人习惯进行排版，文档中字体、文字大小、行间距、段落编号、页边距、落款等参数设置不规范，严重影响到文书的标准性和美观性，以下是文书标准格式要求及日常文档书写注意事项，请各部门在今后工作中严格实行： 一、文件要求 1.文字类采用Word格式排版 2.统计表、一览表等表格统一用Excel格式排版 3.打印材料用纸一般采用国际标准A4型（210mm×297mm），左侧装订。版面方向以纵向为主，横向为辅，可根据实际需要确定 4.各部门的职责、制度、申请、请示等应一事一报，禁止一份行文内同时表述两件工作。 5.各类材料标题应规范书写，明确文件主要内容。 二、文件格式 （一）标题 1.文件标题：标题应由发文机关、发文事由、公文种类三部分组成，黑体小二号字,不加粗，居中，段后空1行。 （二）正文格式 1. 正文字体：四号宋体，在文档中插入表格，单元格内字体用宋体，字号可根据内容自行设定。 2.页边距：上下边距为2.54厘米；左右边距为 3.18厘米。

3.页眉、页脚：页眉为1.5厘米；页脚为1.75厘米； 4.行间距：1.5倍行距。 5.每段前的空格请不要使用空格，应该设置首先缩进2字符 6.年月日表示：全部采用阿拉伯数字表示。 7.文字从左至右横写。 （三）层次序号 (1)一级标题：一、二、三、 (2)二级标题：（一）（二）（三） (3)三级标题：1. 2. 3. (4)四级标题：（1）（2）（3） 注：三个级别的标题所用分隔符号不同，一级标题用顿号“、”例如：一、二、等。二级标题用括号不加顿号，例如：（三）（四）等。三级标题用字符圆点“.”例如：5. 6.等。 （四）、关于落款： 1.对外行文必须落款“湖南环境生物专业技术学院学生会”“校学生会”各部门不得随意使用。 2.各部门文件落款需注明组织名称及部门“湖南环境生物专业技术学院学生会XX部”“校学生会XX部” 3.所有行文落款不得出现“环境生物学院”“湘环学院”“学生会”等表述不全的简称。 4.落款填写至文档末尾右对齐，与前一段间隔2行 5.时间落款：文档中落款时间应以“2016年5月12日”阿拉伯数字

政府公文写作格式规范

政府公文写作格式 一、眉首部分 （一）发文机关标识 平行文和下行文的文件头，发文机关标识上边缘至上页边为62mm，发文机关下边缘至红色反线为28mm。 上行文中，发文机关标识上边缘至版心上边缘为80mm，即与上页边距离为117mm，发文机关下边缘至红色反线为30mm。 发文机关标识使用字体为方正小标宋_GBK，字号不大于22mm×15mm。 （二）份数序号 用阿拉伯数字顶格标识在版心左上角第一行，不能少于2位数。标识为“编号000001” （三）秘密等级和保密期限 用3号黑体字顶格标识在版心右上角第一行，两字中间空一字。如需要加保密期限的，密级与期限间用“★”隔开，密级中则不空字。 （四）紧急程度 用3号黑体字顶格标识在版心右上角第一行，两字中间空一字。如同时标识密级，则标识在右上角第二行。 （五）发文字号 标识在发文机关标识下两行，用3号方正仿宋_GBK字体剧

中排布。年份、序号用阿拉伯数字标识，年份用全称，用六角括号“〔〕”括入。序号不用虚位，不用“第”。发文字号距离红色反线4mm。 （六）签发人 上行文需要标识签发人，平行排列于发文字号右侧，发文字号居左空一字，签发人居右空一字。“签发人”用3号方正仿宋_GBK，后标全角冒号，冒号后用3号方正楷体_GBK标识签发人姓名。多个签发人的，主办单位签发人置于第一行，其他从第二行起排在主办单位签发人下，下移红色反线，最后一个签发人与发文字号在同一行。 二、主体部分 （一）标题 由“发文机关+事由+文种”组成，标识在红色反线下空两行，用2号方正小标宋_GBK，可一行或多行居中排布。 （二）主送机关 在标题下空一行，用3号方正仿宋_GBK字体顶格标识。回行是顶格，最后一个主送机关后面用全角冒号。 （三）正文 主送机关后一行开始，每段段首空两字，回行顶格。公文中的数字、年份用阿拉伯数字，不能回行，阿拉伯数字：用3号Times New Roman。正文用3号方正仿宋_GBK，小标题按照如下排版要求进行排版：

tabularx宏包中改变弹性列的宽度

tabularx宏包中改变弹性列的宽度\hsize 分类：latex 2012-03-07 21:54 12人阅读评论(0) 收藏编辑删除 \documentclass{article} \usepackage{amsmath} \usepackage{amssymb} \usepackage{latexsym} \usepackage{CJK} \usepackage{tabularx} \usepackage{array} \newcommand{\PreserveBackslash}[1]{\let \temp =\\#1 \let \\ = \temp} \newcolumntype{C}[1]{>{\PreserveBackslash\centering}p{#1}} \newcolumntype{R}[1]{>{\PreserveBackslash\raggedleft}p{#1}} \newcolumntype{L}[1]{>{\PreserveBackslash\raggedright}p{#1}} \begin{document} \begin{CJK*}{GBK}{song} \CJKtilde \begin{tabularx}{10.5cm}{|p{3cm} |>{\setlength{\hsize}{.5\hsize}\centering}X |>{\setlength{\hsize}{1.5\hsize}}X|} %\hsize是自动计算的列宽度，上面{.5\hsize}与{1.5\hsize}中的\hsize前的数字加起来必须等于表格的弹性列数量。对于本例，弹性列有2列，所以“.5+1.5=2”正确。 %共3列，总列宽为10.5cm。第1列列宽为3cm，第3列的列宽是第2列列宽的3倍，其宽度自动计算。第2列文字左右居中对齐。注意：\multicolum命令不能跨越X列。 \hline 聪明的鱼儿在咬钩前常常排祠再三& 这是因为它们要荆断食物是否安全&知果它们认为有危险\\ \hline 它们枕不会吃& 如果它们判定没有危险& 它们就食吞钩\\ \hline 一眼识破诱饵的危险，却又不由自主地去吞钩的& 那才正是人的心理而不是鱼的心理& 是人的愚合而不是鱼的恳奋\\

2-1论文写作要求与格式规范(2009年修订)

广州中医药大学研究生学位论文基本要求与写作规范 为了进一步提高学位工作水平和学位论文质量，保证我校学位论文在结构和格式上的规范与统一，特做如下规定： 一、学位论文基本要求 （一）科学学位硕士论文要求 1.论文的基本科学论点、结论，应在中医药学术上和中医药科学技术上具有一定的理论意义和实践价值。 2.论文所涉及的内容，应反映出作者具有坚实的基础理论和系统的专门知识。 3.实验设计和方法比较先进，并能掌握本研究课题的研究方法和技能。 4.对所研究的课题有新的见解。 5.在导师指导下研究生独立完成。 6.论文字数一般不少于3万字，中、英文摘要1000字左右。 （二）临床专业学位硕士论文要求 临床医学硕士专业学位申请者在临床科研能力训练中学会文献检索、收集资料、数据处理等科学研究的基本方法,培养临床思维能力与分析能力,完成学位论文。 1.学位论文包括病例分析报告及文献综述。 2.学位论文应紧密结合中医临床或中西结合临床实际，以总结临床实践经验为主。 3.学位论文应表明申请人已经掌握临床科学研究的基本方法。 4.论文字数一般不少于15000字，中、英文摘要1000字左右。 （三）科学学位博士论文要求 1.研究的课题应在中医药学术上具有较大的理论意义和实践价值。 2.论文所涉及的内容应反映作者具有坚实宽广的理论基础和系统深入的专门知识，并表明作者具有独立从事科学研究工作的能力。 3.实验设计和方法在国内同类研究中属先进水平，并能独立掌握本研究课题的研究方法和技能。

4.对本研究课题有创造性见解，并取得显著的科研成果。 5.学位论文必须是作者本人独立完成，与他人合作的只能提出本人完成的部分。 6.论文字数不少于5万字，中、英摘要3000字；详细中文摘要(单行本)1万字左右。 （四）临床专业学位博士论文要求 1.要求论文课题紧密结合中医临床或中西结合临床实际，研究结果对临床工作具有一定的应用价值。 2.论文表明研究生具有运用所学知识解决临床实际问题和从事临床科学研究的能力。 3.论文字数一般不少于3万字，中、英文摘要2000字；详细中文摘要(单行本)5000字左右。 二、学位论文的格式要求 （一）学位论文的组成 博士、硕士学位论文一般应由以下几部分组成，依次为：1.论文封面；2. 原创性声明及关于学位论文使用授权的声明；3.中文摘要；4.英文摘要；5.目录； 6.引言； 7.论文正文； 8.结语； 9.参考文献；10.附录；11.致谢。 1.论文封面：采用研究生处统一设计的封面。论文题目应以恰当、简明、引人注目的词语概括论文中最主要的内容。避免使用不常见的缩略词、缩写字，题名一般不超过30个汉字。论文封面“指导教师”栏只写入学当年招生简章注明、经正式遴选的指导教师1人，协助导师名字不得出现在论文封面。 2．原创性声明及关于学位论文使用授权的声明（后附）。 3.中文摘要：要说明研究工作目的、方法、成果和结论。并写出论文关键词3～5个。 4.英文摘要：应有题目、专业名称、研究生姓名和指导教师姓名，内容与中文提要一致，语句要通顺，语法正确。并列出与中文对应的论文关键词3～5个。 5.目录：将论文各组成部分(1～3级)标题依次列出，标题应简明扼要，逐项标明页码，目录各级标题对齐排。 6.引言：在论文正文之前，简要说明研究工作的目的、范围、相关领域前人所做的工作和研究空白，本研究理论基础、研究方法、预期结果和意义。应言简

公文写作毕业论文写作要求和格式规范

（公文写作）毕业论文写作要求和格式规范

中国农业大学继续教育学院 毕业论文写作要求和格式规范 壹、写作要求 （壹）文体 毕业论文文体类型壹般分为：试验论文、专题论文、调查方案、文献综述、个案评述、计算设计等。学生根据自己的实际情况，能够选择适合的文体写作。 （二）文风 符合科研论文写作的基本要求：科学性、创造性、逻辑性、实用性、可读性、规范性等。写作态度要严肃认真，论证主题应有壹定理论或应用价值；立论应科学正确，论据应充实可靠，结构层次应清晰合理，推理论证应逻辑严密。行文应简练，文笔应通顺，文字应朴实，撰写应规范，要求使用科研论文特有的科学语言。 （三）论文结构和排列顺序 毕业论文，壹般由封面、独创性声明及版权授权书、摘要、目录、正文、后记、参考文献、附录等部分组成且按前后顺序排列。 1．封面：毕业论文（设计）封面（见文件5）具体要求如下： （1）论文题目应能概括论文的主要内容，切题、简洁，不超过30字，可分俩行排列； （2）层次：高起本，专升本，高起专； （3）专业名称：现开设园林、农林经济管理、会计学、工商管理等专业，应按照标准表述填写； （4）密级：涉密论文注明相应保密年限； （5）日期：毕业论文完成时间。 2．独创性声明和关于论文使用授权的说明：（略）。

3．摘要：论文摘要的字数壹般为300字左右。摘要是对论文的内容不加注释和评论的简短陈述，是文章内容的高度概括。主要内容包括：该项研究工作的内容、目的及其重要性；所使用的实验方法；总结研究成果，突出作者的新见解；研究结论及其意义。摘要中不列举例证，不描述研究过程，不做自我评价。 论文摘要后另起壹行注明本文的关键词，关键词是供检索用的主题词条，应采用能够覆盖论文内容的通用专业术语，符合学科分类，壹般为3～5个，按照词条的外延层次从大到小排列。 4．目录（目录示例见附件3）：独立成页，包括论文中的壹级、二级标题、后记、参考文献、和附录以及各项所于的页码。 5．正文：包括前言、论文主体和结论 前言：为正文第壹部分内容，简单介绍本项研究的背景和国内外研究成果、研究现状，明确研究目的、意义以及要解决的问题。 论文主体：是全文的核心部分，于正文中应将调查、研究中所得的材料和数据加工整理和分析研究，提出论点，突出创新。内容可根据学科特点和研究内容的性质而不同。壹般包括：理论分析、计算方法、实验装置和测试方法、对实验结果或调研结果的分析和讨论，本研究方法和已有研究方法的比较等方面。内容要求论点正确，推理严谨，数据可靠，文字精炼，条理分明，重点突出。 结论：为正文最后壹部分，是对主要成果的归纳和总结，要突出创新点，且以简练的文字对所做的主要工作进行评价。 6．后记：对整个毕业论文工作进行简单的回顾总结，对给予毕业论文工作提供帮助的组织或个人表示感谢。内容应尽量简单明了，壹般为200字左右。 7．参考文献：是论文不可或缺的组成部分。它既可反映毕业论文工作中取材广博程度，又可反映文稿的科学依据和作者尊重他人研究成果的严肃态度，仍能够向读者提供有关

配合前面的ntheorem宏包产生各种定理结构

%=== 配合前面的ntheorem宏包产生各种定理结构,重定义一些正文相关标题===% \theoremstyle{plain} \theoremheaderfont{\normalfont\rmfamily\CJKfamily{hei}} \theorembodyfont{\normalfont\rm\CJKfamily{song}} \theoremindent0em \theoremseparator{\hspace{1em}} \theoremnumbering{arabic} %\theoremsymbol{} %定理结束时自动添加的标志 \newtheorem{definition}{\hspace{2em}定义}[chapter] %\newtheorem{definition}{\hei 定义}[section] %!!!注意当section为中国数字时，[sction]不可用！ \newtheorem{proposition}{\hspace{2em}命题}[chapter] \newtheorem{property}{\hspace{2em}性质}[chapter] \newtheorem{lemma}{\hspace{2em}引理}[chapter] %\newtheorem{lemma}[definition]{引理} \newtheorem{theorem}{\hspace{2em}定理}[chapter] \newtheorem{axiom}{\hspace{2em}公理}[chapter] \newtheorem{corollary}{\hspace{2em}推论}[chapter] \newtheorem{exercise}{\hspace{2em}习题}[chapter] \theoremsymbol{$\blacksquare$} \newtheorem{example}{\hspace{2em}例}[chapter] \theoremstyle{nonumberplain} \theoremheaderfont{\CJKfamily{hei}\rmfamily} \theorembodyfont{\normalfont \rm \CJKfamily{song}} \theoremindent0em \theoremseparator{\hspace{1em}} \theoremsymbol{$\blacksquare$} \newtheorem{proof}{\hspace{2em}证明} \usepackage{amsmath}%数学 \usepackage[amsmath,thmmarks,hyperref]{ntheorem} \theoremstyle{break} \newtheorem{example}{Example}[section]

论文写作格式规范与要求(完整资料).doc

【最新整理，下载后即可编辑】 广东工业大学成人高等教育 本科生毕业论文格式规范（摘录整理） 一、毕业论文完成后应提交的资料 最终提交的毕业论文资料应由以下部分构成： （一）毕业论文任务书（一式两份，与论文正稿装订在一起）（二）毕业论文考核评议表（一式三份，学生填写表头后发电子版给老师） （三）毕业论文答辩记录（一份, 学生填写表头后打印出来，答辩时使用） （四）毕业论文正稿（一式两份，与论文任务书装订在一起），包括以下内容： 1、封面 2、论文任务书 3、中、英文摘要（先中文摘要，后英文摘要，分开两页排版） 4、目录 5、正文（包括：绪论、正文主体、结论） 6、参考文献 7、致谢 8、附录（如果有的话） （五）论文任务书和论文正稿的光盘

二、毕业论文资料的填写与装订 毕业论文须用计算机打印，一律使用A4打印纸，单面打印。 毕业论文任务书、毕业论文考核评议表、毕业论文正稿、答辩纪录纸须用计算机打印，一律使用A4打印纸。答辩提问记录一律用黑色或蓝黑色墨水手写，要求字体工整，卷面整洁；任务书由指导教师填写并签字，经主管院领导签字后发出。 毕业论文使用统一的封面，资料装订顺序为：毕业论文封面、论文任务书、考核评议表、答辩记录、中文摘要、英文摘要、目录、正文、参考文献、致谢、附录（如果有的话）。论文封面要求用A3纸包边。 三、毕业论文撰写的内容与要求 一份完整的毕业论文正稿应包括以下几个方面： （一）封面（见封面模版） （二）论文题目（填写在封面上，题目使用2号隶书，写作格式见封面模版） 题目应简短、明确，主标题不宜超过20字；可以设副标题。（三）论文摘要（写作格式要求见《摘要、绪论、结论、参考文献写作式样》P1～P2） 1、中文“摘要”字体居中，独占一页

Groff 应用

使用Groff 生成独立于设备的文档开始之前 了解本教程中包含的内容和如何最好地利用本教程，以及在使用本教程的过程中您需要完成的工作。 关于本教程 本教程提供了使用Groff(GNU Troff)文档准备系统的简介。其中介绍了这个系统的工作原理、如何使用Groff命令语言为其编写输入、以及如何从该输入生成各种格式的独立于设备的排版文档。 本教程所涉及的主题包括： 文档准备过程 输入文件格式 语言语法 基本的格式化操作 生成输出 目标 本教程的主要目标是介绍Groff，一种用于文档准备的开放源码系统。如果您需要在应用程序中构建文档或帮助文件、或为客户和内部使用生成任何类型的打印或屏幕文档(如订单列表、故障单、收据或报表)，那么本教程将向您介绍如何开始使用Groff以实现这些任务。 在学习了本教程之后，您应该完全了解Groff的基本知识，包括如何编写和处理基本的Groff输入文件、以及如何从这些文件生成各种输出。

先决条件 本教程的目标读者是入门级到中级水平的UNIX?开发人员和管理员。您应 该对使用UNIX命令行Shell和文本编辑器有基本的了解。 系统要求 要运行本教程中的示例，您需要访问运行UNIX操作系统并安装了下面这些软件的计算机(请参见本教程的参考资料部分以获取相关链接)： Groff。Groff分发版中包括groff前端工具、troff后端排版引擎和本教 程中使用的各种附属工具。 自由软件基金会将Groff作为其GNU Project中的一部分进行了发布，所 发布的源代码符合GNU通用公共许可证(GPL)并得到了广泛的移植，几乎对于所有的UNIX操作系统、以及非UNIX操作系统(如Microsoft?Windows?)都有相应 的可用版本。 在撰写本教程时，最新的Groff发布版是Version 1.19.2，对于学习本教 程而言，您至少需要Groff Version 1.17。 gxditview。从Version 1.19.2开始，Groff中包含了这个工具，而在以 前的版本中，对其进行了单独的发布。 PostScript Previewer，如ghostview、gv或showpage。 如果您是从源代码安装Groff，那么请参考Groff源代码分发版中的自述 文件，其中列举了所需的任何额外的软件，而在编译和安装Groff时可能需要 使用这些软件。 介绍Groff 用户通常在字处理软件、桌面发布套件和文本布局应用程序等应用程序环 境中创建文档，而在这些环境中，最终将对文档进行打印或导出为另一种格式。整个文档准备过程，从创建到最后的输出，都发生在单个应用程序中。文档通

TeX 使用指南(常见问题)

TeX 使用指南 常见问题（一） 1.\makeatletter 和\makeatother 的用法？ 答：如果需要借助于内部有\@字符的命令，如\@addtoreset，就需要借助于另两个命令 \makeatletter, \makeatother。 下面给出使用范例，用它可以实现公式编号与节号的关联。 \begin{verbatim} \documentclass{article} ... \makeatletter % '@' is now a normal "letter" for TeX \renewcommand\theequation{\thesection.\arabic{equation}} \@addtoreset{equation}{section} \makeatother % '@' is restored as a "non-letter" character for TeX \begin{document} ... \end{verbatim} 2.比较一下CCT与CJK的优缺点？ 答：根据王磊的经验，CJK 比CCT 的优越之处有以下几点： 1）字体定义采用LaTeX NFSS 标准，生成的DVI 文件不必像CCT 那样需要用patchdvi 处理后才能预览和打印。而且一般GB 编码的文件也不必进行预处理就可直接用latex 编译。2）可使用多种TrueType 字体和Type1 字体，生成的PDF 文件更清楚、漂亮。 3）能同时在文章中使用多种编码的文字，如中文简体、繁体、日文、韩文等。 当然，CCT 在一些细节上，如字体可用中文字号，字距、段首缩进等。毕竟CJK 是老外作的吗。 谈到MikTeX 和fpTeX, 应该说谈不上谁好谁坏，主要看个人的喜好了。MikTeX 比较小，不如fpTeX 里提供的TeX 工具，宏包全，但一般的情况也足够了。而且Yap 比windvi 要好用。fpTeX 是teTeX 的Windows 实现，可以说各种TeX 的有关软件基本上都包括在内。 3.中文套装中如何加入新的.cls文件？ 答：放在tex文件的同一目录下，或者miktex/localtexmf/tex/latex/下的某个子目录下，可以自己建一个。 4.怎样象第几章一样，将参考文献也加到目录？ 答：在参考文献部分加入 \addcontentsline{toc}{chapter}{参考文献}

论文的写作格式及规范

论文的写作格式及规范

附件9： 科学技术论文的写作格式及规范 用非公知公用的缩写词、字符、代号，尽量不出现数学式和化学式。 2作者署名和工作单位标引和检索，根据国家有关标准、数据规范为了提高技师、高级技师论文的学术质量，实现论文写的科学化、程序化和规范化，以利于科技信息的传递和科技情报的作评定工作，特制定本技术论文的写作格式及规范。望各位学员在注重科学研究的同时，做好科技论文撰写规范化工作。 1 题名 题名应以简明、确切的词语反映文章中最重要的特定内容，要符合编制题录、索引和检索的有关原则，并有助于选定关键词。 中文题名一般不宜超过20 个字，必要时可加副题名。英文题名应与中文题名含义一致。 题名应避免使作者署名是文责自负和拥有著作权的标志。作者姓名署于题名下方，团体作者的执笔人也可标注于篇首页地脚或文末，简讯等短文的作者可标注于文末。 英文摘要中的中国人名和地名应采用《中国人名汉语拼音字母拼写法》的有关规定；人名姓前名后分写，姓、名的首字母大写，名字中间不加连字符；地名中的专名和通名分写，每分写部分的首字母大写。 作者应标明其工作单位全称、省及城市名、邮编( 如“齐齐哈尔电业局黑龙江省齐齐哈尔市(161000) ”)，同时，在篇首页地脚标注第一作者的作者简介，内容包括姓名，姓别，出生年月，学位，职称，研究成果及方向。

3摘要 论文都应有摘要（3000 字以下的文章可以略去）。摘要的：写作应符合GB6447-86的规定。摘要的内容包括研究的目的、方法、结果和结论。一般应写成报道性文摘，也可以写成指示性或报道-指示性文摘。摘要应具有独立性和自明性，应是一篇完整的短文。一般不分段，不用图表和非公知公用的符号或术语，不得引用图、表、公式和参考文献的序号。中文摘要的篇幅：报道性的300字左右，指示性的100 字左右，报道指示性的200字左右。英文摘要一般与中文摘要内容相对应。 4关键词 关键词是为了便于作文献索引和检索而选取的能反映论文主题概念的词或词组，一般每篇文章标注3?8个。关键词应尽量从《汉语主题词表》等词表中选用规范词——叙词。未被词表收录的新学科、新技术中的重要术语和地区、人物、文献、产品及重要数据名称，也可作为关键词标出。中、英文关键词应一一对应。 5引言 引言的内容可包括研究的目的、意义、主要方法、范围和背景等。 应开门见山，言简意赅，不要与摘要雷同或成为摘要的注释，避免公式推导和一般性的方法介绍。引言的序号可以不写，也可以写为“ 0”，不写序号时“引言”二字可以省略。 6论文的正文部分 论文的正文部分系指引言之后，结论之前的部分，是论文的核心， 应按GB7713--87 的规定格式编写。 6.1层次标题