A Hierarchical Artificial Neural Network Model for Giemsa-Stained Human Chromosome Classification JONGMAN CHO 1 1 Department of Biomedical Engineering, Inje University, Gimhae, 621-749, KOREA minerva@ieeeorg Abstract: -This paper proposes an improved two-step classification scheme for Giemsa-stained human chromosomes using a hierarchical multilayer neural network with an error back-propagation training algorithm In the first step, the Group (GC), a two-layer neural network, classifies chromosomes into seven groups based on their morphological features, such as relative length, relative area, and the centromeric index and 80 density values In the second step, seven Subgroup s (SCs), which are also two-layer neural networks, classify the chromosomes in each group into 24 subgroups based on the same features used in the first step The optimal parameters for the GC and SCs, including the number of processing elements in the hidden layer, were determined experimentally The optimized GC and SCs were trained using a training dataset and tested using the same test dataset used in a previous study [3] The overall classification error rate decreased to 59% using the two-step classification scheme, which is better than the result in the previous study, which used a single-step multilayer neural network as a classifier and achieved a 652% classification error rate [3] This paper shows that the classification accuracy for Giemsa-stained human chromosomes can be improved using a two-step classification scheme rather than a single-step classification Key-Words: - Giemsa-stained human chromosome, classification, hierarchical multilayer neural network 1 Introduction Human chromosome analysis is an essential task in cytogenetics, especially in prenatal screening, genetic syndrome diagnosis, cancer pathology research, and environmentally induced mutagen dosimetry [1] Cells used for chromosome analysis are taken mostly from amniotic fluid or blood samples The stage at which chromosomes are most suitable for analysis is the metaphase in cellular division One of the objectives of chromosome analysis is to create a karyotype, which is a layout of paired chromosome images organized by decreasing size The karyotype is obtained after cell culture, preparing slides, selecting the best chromosome image, analysis, and classification However, chromosome analysis and karyotyping are still performed manually in most cytogenetics laboratories in a repetitive, time-consuming, and therefore expensive process [2] Efforts to develop automated chromosome classification techniques have been made over the last 20 years The use of features rather than the picture itself makes the classification procedure easier and faster All the efforts to automate chromosome analysis have had limited success and poor classification results compared to those of a trained cytotechnician [1] [6] An artificial neural network (ANN) can overcome most of these limitations because it allows the application of expert knowledge and experience through network training In addition, it is suitable for automatic chromosome classification because human chromosome images have nonlinear properties [3] Finally, it is well-known that the problems best solved by neural networks are those that humans do well, and the classification of chromosomes is one of them [2] This paper proposes a hierarchical multilayer neural network (HMLNN) as a human chromosome classifier The first-step neural network classifies chromosome data into seven major groups based on morphological features, such as relative length and relative area, and the centromeric index and 80 density values The second-step neural network classifies each of these seven major groups into subgroups 2 Methodology 21 Chromosome Database The suggested methodology was applied to the Edinburgh database, provided by Dr Piper and exported from the Medical Research Council Human Genetics Unit, Edinburgh, UK The chromosomes in this database were photographed and digitized individually from photographic negatives on a grid that paralleled the chromosome axis, using a microdensitometer Each image consisted of a rectangular array of pixels comprising the chromosome bordered by the adjacent background ISSN: 1790-2769 Page 44 ISBN: 978-960-6766-32-9
22 Experimental Tools In this study, Microsoft Visual C++ Net was used to implement the application program for feature extraction and NeuroShell 20 (The Ward Systems Group) was used to train and test the proposed neural networks 23 Feature Extraction 231 Medial Axis The two chromatids of a metaphase chromosome are located symmetrically, but most chromosomes are not exactly straight Piper reviewed several ways to obtain the medial axis of a chromosome, which is the virtual symmetric line of the two chromatids of a chromosome However, we used a different method to obtain the medial axis, which was a modification of the thinning algorithm of Gonzalez [5] The medial axis of a chromosome is obtained by applying the medial axis transformation (MAT) algorithm [3] In addition, we used a thinning algorithm that iteratively deletes edge points of a region, subject to the constraints that the deletion of these points does not remove end points, does not break connectedness, and does not cause excessive erosion of the region [4] After applying the MAT, the skeletonized line was extended to the boundary and the extended medial axis of the chromosomes was obtained, as shown in Fig 1(b) 232 Relative Length (RL) One important morphological feature used to identify a chromosome is a length characteristic However, the length must be normalized before identifying a chromosome because it varies according to the phase of cellular division After determining the medial axis, the Euclidean distance is commonly used to obtain the distance between two points However, this method was not used here because most chromosomes are not exactly straight Instead, the length of each chromosome could be calculated by counting the number of pixels along the medial line The relative length of the i-th chromosome (l ri ) can be obtained by normalizing the medial axis length using the following equation: l i l ri =, (1) lt where l i (i = 1, 2,, 24) is the length of the i-th chromosome and l t is the total length of all 46 chromosomes in one cell 233 Relative Area (RA) The relative area is another important morphological feature used to identify a chromosome Like the relative length, the relative area must be normalized The relative area of the i-th chromosome (A ri ) can be obtained by counting the pixels in the chromosome body and by normalizing the area using the following equation: A i A ri =, (2) At where A i (i = 1, 2,, 24) is the area of the i-th chromosome and A t is the total area of all 46 chromosomes in one cell 234 Centromeric Index (CI) The centromeric index is the ratio of the length of the short arm to the total length of a chromosome It is also an important morphological feature used to identify the chromosome The CI, which indicates the location of the centromere on the chromosome, is obtained from the shape profile, as shown Fig 1(c) It is a one-dimensional graph of the chromosome computed as a sequence of points along the possible curved chromosome medial axis The shape profile for a chromosome is obtained by measuring the width along a transverse line, perpendicular to the tangent of the medial axis, and centered a unit distance along the medial axis The CI is the ratio of the length of the short arm to the total length of one chromosome, as shown in Fig 2 It is another important morphological feature used to identify the chromosome and can be calculated using the following equation: l si C i =, (3) li where C i is the CI of the i-th chromosome, l si (i = 1, 2,, 24) is the length of the short arm, and l i is the total length of the i-th chromosome 235 Normalized Density Profile It is difficult to determine the location of the centromere and the centromeric indices for telocentric chromosomes in groups D (no 13, 14, and 15) and G (no 21, 22, and Y) A Giemsa-stained human chromosome has a sequence of bands that are perpendicular to the medial axis of the chromosome This feature is important for identifying chromosomes because each chromosome has its own band pattern ISSN: 1790-2769 Page 45 ISBN: 978-960-6766-32-9
Length of the short arm (a) Centromere Total chromosome length (b) (c) (d) Fig 1 Feature extraction for a Giemsa-stained no 1 human chromosome: (a) raw 256-gray level image; (b) binary image and medial axis obtained from the raw image; (c) the shape profile is obtained from the binary image, and the location of the centromere is determined using the shape profile (the vertical line indicates the location of the centromere); (d) the density profile before normalization The density profile (DP) is a one-dimensional graph of the banding pattern of a chromosome computed at a sequence of points along the curved chromosome medial axis The DP of a chromosome is obtained from measurements made along a transverse line, perpendicular to the tangent of the medial axis, and centered at a unit distance along the medial axis, as shown in Fig 1(d) The density value (I k ) for every point on the medial axis results from summing the properties of points spaced a unit distance apart along each transverse line and can be calculated as follows: m 1 = I k d( k, j) ( k = 0,1, K, n 1), (4) j= 0 where n is the number of pixels on the medial axis, m is the number of pixels on the line perpendicular to the tangent to the medial axis, and d(k, j) is the pixel value on the line perpendicular to the tangent to the medial axis Fig 2 The centromere and short arm of chromosome no 2 The Centromeric Index (CI) is the ratio of the length of the short arm to the total length of the chromosome To reduce the variation in the density values due to different cell culture conditions and nonhomogeneous microscope illumination conditions, this DP was normalized three times using three different methods First, this profile is normalized in the direction of the perpendicular to the medial axis of the chromosome to reduce the variation in the width of the chromosome as follows: I k dw( k) = ( k = 0,1, K, n 1), (5) w( k) where w(k) is the width of the k-th point on a chromosome and d w (k) is the density value normalized in the width direction Next, a histogram equalization is applied to d w (k) to reduce the effect of nonhomogeneous illumination as follows: d w ( k) d wmin d N ( k) = ( k = 0,1, K, n 1), (6) d d wmax wmin where d N (k) is the histogram equalized density value, d wmin is the minimum density value, and d wmax is the maximum density value of that chromosome Finally, d N (k) is normalized along the medial axis to obtain the same number of density values regardless of the chromosome length In this study, 80 density values were obtained for each chromosome based on the results of a previous study [9] 236 Chromosome Database Two datasets were prepared for training and testing the HMLNN proposed here The training dataset consisted ISSN: 1790-2769 Page 46 ISBN: 978-960-6766-32-9
of 460 input patterns extracted from 460 chromosomes, which were obtained from cells from 10 different people In this training dataset, one input pattern consisted of 83 features extracted from one chromosome: the relative length, l ri, relative area, A ri, centromeric index, C I, and 80 density values The test dataset has the same number of data patterns as the training dataset, but the features were extracted from chromosomes that were not used to prepare the training dataset In this study, the test dataset was obtained from chromosome data used in a previous study [3] for comparison 237 Chromosome s This study examined the classification error rate of HMLNN as a chromosome classifier HMLNN consists of hierarchically cascaded classifiers: one Group (GC) and seven Subgroup s (SCs), as shown in Fig 3 The GC and SCs are a two-layer neural network with an error back-propagation training algorithm Chromosome Features Group Group (GC) Group A Group B Group C Group D Group E Group F Group G Fig 3 Architecture of the hierarchical multilayer neural network (HMLNN) 1 2 3 21 22 Y Subgroup s (SCs) Group (GC) This classifier divides the chromosomes into seven groups (Groups A G) The number of input nodes and processing elements (PEs) in the output layer of the GC was fixed to 83 and 7, respectively, based on the number of features extracted from one chromosome and the number of groups The number of PEs in the hidden layer of the GC was determined experimentally Subgroup s (SCs) Seven neural networks were designed and used to classify the chromosomes in each group as individual chromosomes The number of input nodes of the neural network for each SC was fixed to 83, the same number as in the GC, but the numbers of PEs in the output layer and hidden layer differed, depending on the groups and values determined experimentally 3 Results 31 Number of Processing Elements in the GC and SCs The optimal numbers of PEs in the hidden layer for the GC and SCs were determined separately in experiments The optimal number of PEs in the hidden layer of the GC, which gives the minimum classification error, was determined by repeating the experiment for various numbers of PEs in the hidden layer As shown in Table 1, a GC with 35 PEs in its hidden layer had the smallest classification error The number of input nodes of SCs was also fixed at 83, but the number of PEs in the output layer depended on the number of chromosomes in that group, as shown in Table 2 The optimal number of PEs in the hidden layer of the SCs was also determined by repeating the experiment for various numbers of PEs in the hidden layer Table 3 shows the selected numbers of PEs for the seven SCs 32 The Group (GC) Classification After the number of PEs in the hidden layer was determined, the two-layer neural network for the GC was built, trained 2,000 times using the training dataset, and tested using the test dataset, which was not used for the training phase The resulting GC classification is shown in Table 4 In all, the GC misclassified 16 of 460 chromosomes 33 The Subgroup (SC) Classification The seven SCs configured using the parameters shown in Tables 2 and 3 were trained 2,000 times using the training data that were classified correctly using the GC After training, each SC was tested using the same test dataset used by the GC in its test phase Only the chromosome data correctly classified by the GC were classified by the ISSN: 1790-2769 Page 47 ISBN: 978-960-6766-32-9
seven different SCs The SCs misclassified 11 chromosomes, as shown in Table 5 Table 1 Classification error depending on the number of PEs in the hidden layer of the Group (GC) mber of PEs in the hidden l Number of misclassified chromosomes 20 22 25 24 30 25 35 16 40 27 45 27 50 21 Table 2 Number of PEs in the output layer of the Subgroup s (SCs) Group Number of PEs in the output layer A 3 B 2 C 8 D 3 E 3 F 2 G 3 Table 3 Number of PEs in the hidden layer of the Subgroup s (SCs) Subgroup s (SCs) Number of PEs in the hidden layer SC-A 7 SC-B 7 SC-C 11 SC-D 15 SC-E 11 SC-F 7 SC-G 11 Table 4 Result of the Group (GC) classification cation using the trained neural network, Group Classif fi-cation by an A B C D E F G Sum A 59 1 60 B 38 2 40 C 2 1 152 156 D 1 58 1 60 E 1 59 1 60 F 39 1 40 G 5 39 44 Sum 61 39 157 58 60 44 41 460 for Giemsa-stained human chromosomes In the first step, the Group (GC) classified Giemsa-stained human chromosomes into seven groups based on their morphological features, such as the relative length and relative area, and the centromeric index and 80 density values In the second step, seven Subgroup s (SCs) separated the chromosomes that had been classified correctly by the GC into 24 categories The numbers of processing elements (PEs) in the hidden layers of the GC and SCs were determined experimentally The GC and SCs were built using NeuroShell with parameters determined experimentally and trained using a test dataset prepared from 460 chromosomes After 2,000 training runs, the GC and SCs were tested using a test dataset with the same number of chromosomes, which was not used in the training phase The GC misclassified 16 chromosomes, while the SCs misclassified 11 chromosomes Therefore, 27 of 460 chromosomes were misclassified by the proposed classifier and the overall classification error rate was 587%, which is better than that of a previous study using the same chromosome data, and had a classification error rate of 652% [3] As shown in Tables 4 and 5, there were relatively more classification errors for group G chromosomes: GC misclassified 5 of 44 chromosomes (1136%) and SC-G misclassified 5 of 39 chromosomes (1282%) The group G chromosomes (no 21, 22, and Y) are telocentric, and it was difficult to calculate their centromeric indices (CIs) accurately using the methods proposed here In addition, the chromosomes of this group were the smallest in size of the seven groups, so the 80 density values obtained from the small density profile contained more redundant information than in the other groups The centromere-finding algorithm proposed here must be improved so that it can be applied to telocentric chromosomes and reduce the classification error In addition, new chromosome features that will improve the classification accuracy should be determined 4 Discussion and Conclusion This study proposed a two-step classification scheme using a hierarchical multilayer neural network with an error back-propagation training algorithm as a classifier ISSN: 1790-2769 Page 48 ISBN: 978-960-6766-32-9
Table 5 Classification result using the Subgroup s (SCs) Classification result using the trained neural network A B C D E F G 1 2 3 4 5 6 7 8 9 10 11 12 X 13 14 15 16 17 18 19 20 21 22 Y Sum 1 19 1 20 A 2 19 19 3 20 20 B 4 19 19 5 1 18 19 6 19 19 7 19 19 8 19 1 20 C 9 1 18 1 20 10 20 20 11 1 20 20 12 19 20 X 15 15 13 20 20 D 14 19 19 15 19 19 16 20 20 E 17 18 18 18 20 20 F 19 20 20 20 19 19 21 15 2 1 18 G 22 1 16 17 Y 1 3 4 Sum 19 19 21 20 18 19 19 20 20 20 20 19 16 20 19 19 20 18 20 20 19 17 18 4 444 References: [1] A Carothers and J Piper, Computer-aided classification of human chromosomes, A Review Stat Comput, vol 4, pp 161 171, 1994 [2] B Lerner, Toward a completely automatic neural-network-based human chromosome analysis, IEEE Trans Syst Man Cyber, vol 28, pp 544 552, 1998 [3] J Cho, Chromosome classification using backpropagation neural network, IEEE Med Biol Magn, vol 19, pp 28 33, 2000 [4] J Russell, Genetics, Addison Wesley Longman, New York, 1998 [5] R C Gonzalez and R Woods, Digital Image Processing, Addison-Wesley, MA, 1992 [6] F Groen, T Kate, A Smeulders, and I Young, Human chromosome classification based on local band descriptors, Pattern Recog Lett, vol 9, pp 211 222, 1989 [7] B Lerner, H Guterman, I Dinstein, and Y Romen, Medial axis transform-based features and a neural network for human chromosome classification, Pattern Recog, vol 28, pp 1673 1683, 1995 [8] D Harnden, H Klinger, and J Jensen, An International System for Human Cytogenetic Nomenclature (ISCN 1985): Report of the Standing Committee on Human Cytogenetic Nomenclature, KAEGER, Basel, Switzerland, 1985 [9] S Ryu and J Cho, Feature extraction of Giemsa-stained chromosomes and classification error of a backpropagation neural network-based classifier, Proc SPIE, vol 4668, pp 20 28, 2002 ISSN: 1790-2769 Page 49 ISBN: 978-960-6766-32-9