- Email: [email protected]
Quantifying the Kinetics of Somatic Embryo Development
Copyright © IfAC Computer Applications in Biotechnology, Osaka, Japan, 1998
QUANTIFYING THE KINETICS OF SOMA TIC EMBRYO DEVELOPMENT Chun Zhang and Wei-Shou Hu l
Department of Chemical Engineering and Materials Science, University of Minnesota, Minneapolis. MN 55455-0132
Abstract: Micropropagation by somatic embryogenesis is an important method for propagating elite clones of economically important plants. The quality of the embryo culture is often evaluated by the morphology of the embryo population. We have developed a pattern recognition system to classify developing embryos according to their morphology . The images of embryos of Douglas fir and carrot were processed and the embryo morphological information was converted into a set of digital numbers (Fourier features and size features) through discrete fast Fourier transformation. Neural network classifiers were developed to separate those embryos into different morphological classes based on these Fourier and size features. In addition, the course of each embryo's development can be described as a trajectory in a multi-dimensional feature space spanned by the Fourier and size features . Important insights can be obtained by analyzing the developmental paths of normal and abnormal embryos. This quantitative description of embryo population is therefore a powerful tool for describing the kinetics of embryo development. Copyright © 1998IFAC
Key words: pattern recognition, neural network, somatic embryogenesis, carrot, Douglas fir
1. INTRODUCTION
uniformity is important for the automation of plantlet handling, somatic embryogenesis may offer much advantage over conventional processes.
Vegetative micropropagation via cuttings and shoot culture has been in practice in home gardens as well as commercially for centuries. These procedures have been important tools in the propagation of elite clones for many species. However, such propagation methods are labor intensive and not readily amenable to large-scale operations. As an alternative the possibility of micropropagatioll through somatic embryogenesis for the production of elite or disease free plants has aroused a great deal of interest. Using a suspension culture a large number of embryos can be generated in a relatively small environmentally controlled bioreactor. Mature embryos can be encapsulated in hydrogel as synthetic seeds for transportation and plantation in the field. Plantlets regenerated from micropropagated selected clones of gymnosperms (e.g. Douglas fir, loblolly pine) showed superior field performance over seedlings derived from the same parent (Gupta, Timmis et al. 1991; Timmis 1995). Such somatic embryo derived plantlets often exhibit relatively uniform phenotypes; whereas seedlings from the same parent display variations in traits just as one expects from siblings unless the seeds are derived from completely homozygous lines from haploid plants. Since
Despite the potential advantage of somatic embryogenesis for rnicropropagation, its large-scale application is still rather limited. A major huddle for its wide acceptance is the high cost per cultivar. Contributing to this high cost of production is the low yield and heterogeneity of the culture. Most embryo cultures give rise to a heterogeneous population consisted of embryos at various stages of development, abnormal embryos and callus. This population heterogeneity necessitates a laborious sorting step to select for normal and mature embryos. Furthermore, work on the optimization of culture conditions to improve the yield and the population homogeneity is often hampered by the difficulty in evaluating the results. The assessment of embryo quality has been largely based on the morphology of the embryos. Typically, an operator examines an embryo population sampled from a culture under a microscope, classifies the embryos into different developmental classes and enumerates them. The process is subjective and tedious. Hardly enough embryos are employed to give a high degree of confidence in evaluating the results. More often, definitive comparison on the effect of culture conditions cannot be made unless the results obtained under different conditions are drastically different. As a consequence, kinetic data suitable for process optimization studies are virtually nonexistent. Few
ITo whom all correspondence should be addressed Email: [email protected]
485
kinetic studies on somatic embryo development were followed up by more detailed investigation as it was prohibitively laborious.
contours and eliminate noises generated during the image acquisition. These embryo contours were then transferred to a HP Apollo 9000 series 735 workstation (Hewlett Packard Co .. Chelmsford. MA) for the further morphological and size feature extraction. Each embryo contour was placed on a complex plane. Its center of gravity and the origin of the plane were superimposed. The embryo contour was then normalized in size and its longer axis was aligned with the horizontal axis according to certain criteria. For carrot embryos the contour was then divided into 32 equal arc-length segments to give rise to 32 Fourier coefficients (+1 to +16 and -I to -16) after the discrete and normalized fast Fourier transformation. For Douglas fir. 128 equal arclength segments were used. The magnitude of these Fourier coefficients was defined as the Fourier feature. At the same time. several size features. such as biovolume. perimeter. projection area. and length/width. were also calculated .
In this paper we describe the development of an imaging analysis and pattern recognition system to separate embryo images into different morphological classes objectively and their application in tracking individual embryos' development. The availability of such a pattern recognition system can facilitate kinetic studies of somatic embryo cultivation.
2. MATERIALS AND METHODS 2.1 Cultures and Media Callus and liquid sllspension cultures of carrot (Daucus carota L. cv. Danvers) were initiated from seedling hypocotyls. Cultures were maintained in the modified Murashige and Skoog medium in which an equal weight of glucose was added as a substitute for sucrose. The medium was supplemented with 1.0 mg/L (4 .5 ~M) 2,4-dichlorophenoxyacetic acid (2,4D). Embryo cultures were initiated by removing 2,4D from the medium (Huang. Chi et aI. 1993). Cultures were incubated in the dark in a rotary shaker (New Brunswick. Edison. NI) at 80 rpm and 24± I
A hierarchical decision tree was used for the classification for both Douglas fir and carrot embryos (Figure I and Figure 2). At each node of the decision tree. a group of embryo classes was separated into several subgroups. Each node was considered as one neural network classifier (for details see Materials and Methods). For carrot. the examples fed into the neural classifiers contained 31 Fourier features and 3 size features as inputs . and the corresponding developmental stage as outputs. For Douglas fir, ten size related and nine Fourier features were used .
QC.
Douglas fir (Pseudotsuga menziesii) cells were maintained . as described elsewhere (Taber. Zhang et aI. 1997). The cells from maintenance stage were pretreated in a medium containing high abscisic acid concentration before being transferred to maturation stage. Maturation culture was carried out in 60 x 15 mm Petri dishes . A cellulose membrane with 0.45 flm pore size (Millipore. Bedford. MA) was laid on a 35 x 35 x 5 mm polyester pad placed in the Petri dish. 6 mL maturation medium was added to the dish and absorbed into the polyester pad . This allowed the surface of membrane to be moist but not submerged with medium . Cells from week 3 ABA singulation culture were allowed to settle. The cells were washed with fresh maturation medium twice and then used for inoculum of maturation cultures. 1.2 mL of the cell suspension were transferred to the Petri dish and evenly laid on the surface of the membrane. The Petri dishes were sealed with parafilm and placed in the dark at 24 ±I qc.
Popu lation
l Il
r--G-IOb-u"-Iar--'I
Torpedo
2.2 1mage Analysis and Pattern Recognition The embryo images were obtained by a CCD camera and pre-processed according to the procedures described previously (Chi. Vits et aI. 1994; Vits . Chi et aI. 1994). In brief, the raw image was first binarized by defining a threshold value to discriminate embryos from the background. A thinning algorithm was used to extract embryo
Fig. l. Decison tree of the neural network based pattern recognition system (carrot).
486
embryos into three subgroups. Nodes #3 and #4 were used to separate the oblong+heart and The twins+c1uster groups, respectively . classification agreement between the two methods ranged from 73 to 88% for these four classes.
Table I . testing results of layer #1 of the neural network classifiers (carrot)
Normal Embryo
Abnormal Embryo
Neural Network Classification (%)
K·NN Method
Total Globular Torpedo Callus Others
~ ~
S2 Embryo
S3 Embryo
Globular
126
91.3
0.7
5.6
2.4
Torpedo
182
0.0
90.2
0.5
9.3
Callus
117
6 .0
0.0
93 .2
0.8
Others
403
3.7
6 .9
8.3
81.1
Fig. 2. Decison tree of the neural network based pattern recognition system (Douglas fir) Table 2 testing results of layer #2 of the neural network classifiers (carrot)
3. RESULTS AND DISCUSSION K·NN Method
Neural Network Classification (%)
3.1 Application of neural network for embryo classification.
Total Oblong +Heart
The classification results of each classifier for carrot embryos are shown in Tables I, 2 and 3 for the first. second and third layer classifiers, respectively. In each table, the second column contains the total number of embryos in a particular class as dassi fied previously using k·nearesl neighbor method . The other columns show the results of neural network The diagonal cells show the classification. agreement of the dassification between the two methods. At node #1 , the neural network dassifier separated the population into eight dasses, which were further grouped into four categories· globular. torpedo. callus and "Others" . "Others" contained oblong, heart. secondary, twins and cluster embryos . The agreement of globular, torpedo. callus and "Others" groups hetween the neural network classifier and K·NN method were 91.3 % . 90.2 %, 93.2% and X1.1 % . respectively. Group "Others" . consisting of fi ve embryo classes, was further
Secondary
Twins +Cluster
Oblong +Heart
151
92. 1
7.3
0.6
Secondary
108
19.4
57.4
232
Twins +Cluster
144
07
9.0
90.3
Table 3 Testing results of layer #3 of the neural net work c1assi fiers (carrot) K·NN Method
Total
Twins
Cluster
Twins
67
76. 1
23 .9
Cluster
77
11 .7
88.3
K·NN Method
separated into individual classes using node #2 (Figure I). The neural network classifier at this node was trained by another embryo population which contained only five embryo classes: oblong, heart, twins, cluster and secondary embryos. The embryos were separated into 3 subgroups : oblong+heart. twins+c1uster, and secondary embryos . Node #2 did not give a satisfactory separation of "Others" embryos from node # I into individual classes. Rather, it gave a satisfactory separation of those
Neural Network Classification ('To)
Neural Network Classification (%) Total
Oblong
Heart
Oblong
67
82.9
17 .1
Heart
77
26. 1
73.9
For Douglas fir. the population was first separated into two large groups - normal and abnormal embryos. A 3-layer feed forward neural network
487
classifier was used to accomplish this task. Ten sizerelated features and nine Fourier features (-5 to -2 and + I to +5) that contain the major morphological information of Douglas fir embryos were used as neural network classifier inputs.
Table 6: Comparison of classification of Douglas fir by a human operator and by the neural network classifiers. Human
Table 4 shows the testing results of the classifier at node #1 (Figure 2). The numbers in the second column are the human classification results and the last twO columns show the classification results of the neural network classifier. The classification accuracy for normal embryos was as high as 98 %. Further, the normal embryo group from node #1 was separated into three developmental stages corresponding to young, medium and mature embryos" This was achieved by a second neural network classifier at node #2 of the decision tree . The results are shown in Table 5. Among the 267 normal embryos classified by the operator 66, 136, and 65 fell into SI, S2, and S3 embryos, respectively. 54 out of 66 SI embryos were correctly classified by the neural network classifier, 11 classified as S2 embryos and I as S3 embryo. For S2 and S3 embryos, 118 out of 136 and 55 out of 65 were classified correctly giving a classification accuracy of 87% and 84% for S2 and S3 embryos, respectively. The accuracy of classifications was higher than 80% for all the morphological classes except abnormal embryos. The bottom two rows of Table 6 show the estimated number of embryos in each class and the accuracy of the neural network based pattern recognition system. The accuracy was from 89% for S3 embryos up to 97 % for S2 embryos. Overall , this pattern recognition system gave a good estimation of the embryos number in every class.
Normal Embryo Abnormal Embr;ro
Normal
Abnormal
Embr~o
Embr~o
271
267
4
32
10
22
81 82 83
Total 66 136
65
82
83
54
11 118 10
I
6
o
32
82
83
Abnormal
6 0 2
10 118 10 4
63
142
72
26
93 %
97%
89%
81%
2 2 0 22
12
55 4
ACKNOWLEDGEMENTS
Neural Network Classifier
81
Abnormal Total Accurac;r
81 55
Large variations in carrot embryo development were observed which is consistent with the population heterogeneity seen in batch embryo cultures. The rate at which each embryo progresses in the feature space was measured by a developmental vector. A major potential utility of single embryo tracking is to uncover a set of morphological features at an early embryo stage. that is associated with the eventuaI fate of those individuals. If such correlation could be found, then one can predict whether a particular embryo will become a mature. aborted. or abnormal embryo. even when it is in early stage. The analysis on developmental vectors revealed that embryos with a higher developmental rate during the carly stage of development had a higher chance of reaching the mature stage in a relatively short time . This single embryo tracking method is potentially a valuable tool in developing a correlation between the embryo's morphological features during the early stage of development and its final developmental fate .
Table 5: Testing at node #2 of the neural network classifiers (Douglas fu:1 Human Operator
68 138 65
The Fourier and size features span a multi-dimension feature space in which an embryo is represented by a particular point defined by its Fourier and size features. As the embryo develops, its Fourier and size features change accordingly and its position in the feature space also changes. This results in a series of points which give rise to a trajectory characterizing the developmental journey of this embryo. Connecting each point representing the individual embryos over time gives a trajectory which depicts the embryos' developmental "path" or history. By cultivating carrot embryos on Petri plates and following each individual embryo's development over time. we constructed such "developmental paths" for over 400 embryos.
Neural Network Classifier Total
81 82 83
3.2 Single embryo tracking
Table 4: Testing at node # 1 of the neural network classifiers (Douglas fir). Human Operator
Neural Network Classifiers
This work was supported in part by a grant from the National Science Foundation (BCS 9015817) We acknowledge Dr. Roger Timmis of Weyerhaeuser Corporation for kindly providing Douglas fir cell lines.
12
55
488
REFERENCES Chi, C. M., H. Vits, et al. (1994). Morphological kinetics and distribution in somatic embryo cultures. Biotechnol Bioeng. 44(3) : 368-378. Gupta, P. K., R. Timmis, et al. (1991). Field performance of micropropagated forestry species. In Vitro Cell. Dev. BioI., Plant 27P: 159-164. Huang, L. c., C. M. Chi, et al. (1993). Population and biomass kinetics in fed-batch cultures of Daucus carota L. somatic embryos. Biotechnol. Bioeng. 41(8): 811-818 . Taber. R. P .. C. Zhang. et al. (1997). Kinetics of Douglas fir (Pseudotsuga menziesii) somatic embryo development. Submitted. Timmis. R. (1995). Personal communication . Vits. H .• C.-M. Chi. et al. (1994) . Characterization of patterns in plant somatic embryo cultures : the morphology and development of embryos. AIChEJ . 40 : 1728-1740.
489
QUANTIFYING THE KINETICS OF SOMA TIC EMBRYO DEVELOPMENT Chun Zhang and Wei-Shou Hu l
Department of Chemical Engineering and Materials Science, University of Minnesota, Minneapolis. MN 55455-0132
Abstract: Micropropagation by somatic embryogenesis is an important method for propagating elite clones of economically important plants. The quality of the embryo culture is often evaluated by the morphology of the embryo population. We have developed a pattern recognition system to classify developing embryos according to their morphology . The images of embryos of Douglas fir and carrot were processed and the embryo morphological information was converted into a set of digital numbers (Fourier features and size features) through discrete fast Fourier transformation. Neural network classifiers were developed to separate those embryos into different morphological classes based on these Fourier and size features. In addition, the course of each embryo's development can be described as a trajectory in a multi-dimensional feature space spanned by the Fourier and size features . Important insights can be obtained by analyzing the developmental paths of normal and abnormal embryos. This quantitative description of embryo population is therefore a powerful tool for describing the kinetics of embryo development. Copyright © 1998IFAC
Key words: pattern recognition, neural network, somatic embryogenesis, carrot, Douglas fir
1. INTRODUCTION
uniformity is important for the automation of plantlet handling, somatic embryogenesis may offer much advantage over conventional processes.
Vegetative micropropagation via cuttings and shoot culture has been in practice in home gardens as well as commercially for centuries. These procedures have been important tools in the propagation of elite clones for many species. However, such propagation methods are labor intensive and not readily amenable to large-scale operations. As an alternative the possibility of micropropagatioll through somatic embryogenesis for the production of elite or disease free plants has aroused a great deal of interest. Using a suspension culture a large number of embryos can be generated in a relatively small environmentally controlled bioreactor. Mature embryos can be encapsulated in hydrogel as synthetic seeds for transportation and plantation in the field. Plantlets regenerated from micropropagated selected clones of gymnosperms (e.g. Douglas fir, loblolly pine) showed superior field performance over seedlings derived from the same parent (Gupta, Timmis et al. 1991; Timmis 1995). Such somatic embryo derived plantlets often exhibit relatively uniform phenotypes; whereas seedlings from the same parent display variations in traits just as one expects from siblings unless the seeds are derived from completely homozygous lines from haploid plants. Since
Despite the potential advantage of somatic embryogenesis for rnicropropagation, its large-scale application is still rather limited. A major huddle for its wide acceptance is the high cost per cultivar. Contributing to this high cost of production is the low yield and heterogeneity of the culture. Most embryo cultures give rise to a heterogeneous population consisted of embryos at various stages of development, abnormal embryos and callus. This population heterogeneity necessitates a laborious sorting step to select for normal and mature embryos. Furthermore, work on the optimization of culture conditions to improve the yield and the population homogeneity is often hampered by the difficulty in evaluating the results. The assessment of embryo quality has been largely based on the morphology of the embryos. Typically, an operator examines an embryo population sampled from a culture under a microscope, classifies the embryos into different developmental classes and enumerates them. The process is subjective and tedious. Hardly enough embryos are employed to give a high degree of confidence in evaluating the results. More often, definitive comparison on the effect of culture conditions cannot be made unless the results obtained under different conditions are drastically different. As a consequence, kinetic data suitable for process optimization studies are virtually nonexistent. Few
ITo whom all correspondence should be addressed Email: [email protected]
485
kinetic studies on somatic embryo development were followed up by more detailed investigation as it was prohibitively laborious.
contours and eliminate noises generated during the image acquisition. These embryo contours were then transferred to a HP Apollo 9000 series 735 workstation (Hewlett Packard Co .. Chelmsford. MA) for the further morphological and size feature extraction. Each embryo contour was placed on a complex plane. Its center of gravity and the origin of the plane were superimposed. The embryo contour was then normalized in size and its longer axis was aligned with the horizontal axis according to certain criteria. For carrot embryos the contour was then divided into 32 equal arc-length segments to give rise to 32 Fourier coefficients (+1 to +16 and -I to -16) after the discrete and normalized fast Fourier transformation. For Douglas fir. 128 equal arclength segments were used. The magnitude of these Fourier coefficients was defined as the Fourier feature. At the same time. several size features. such as biovolume. perimeter. projection area. and length/width. were also calculated .
In this paper we describe the development of an imaging analysis and pattern recognition system to separate embryo images into different morphological classes objectively and their application in tracking individual embryos' development. The availability of such a pattern recognition system can facilitate kinetic studies of somatic embryo cultivation.
2. MATERIALS AND METHODS 2.1 Cultures and Media Callus and liquid sllspension cultures of carrot (Daucus carota L. cv. Danvers) were initiated from seedling hypocotyls. Cultures were maintained in the modified Murashige and Skoog medium in which an equal weight of glucose was added as a substitute for sucrose. The medium was supplemented with 1.0 mg/L (4 .5 ~M) 2,4-dichlorophenoxyacetic acid (2,4D). Embryo cultures were initiated by removing 2,4D from the medium (Huang. Chi et aI. 1993). Cultures were incubated in the dark in a rotary shaker (New Brunswick. Edison. NI) at 80 rpm and 24± I
A hierarchical decision tree was used for the classification for both Douglas fir and carrot embryos (Figure I and Figure 2). At each node of the decision tree. a group of embryo classes was separated into several subgroups. Each node was considered as one neural network classifier (for details see Materials and Methods). For carrot. the examples fed into the neural classifiers contained 31 Fourier features and 3 size features as inputs . and the corresponding developmental stage as outputs. For Douglas fir, ten size related and nine Fourier features were used .
QC.
Douglas fir (Pseudotsuga menziesii) cells were maintained . as described elsewhere (Taber. Zhang et aI. 1997). The cells from maintenance stage were pretreated in a medium containing high abscisic acid concentration before being transferred to maturation stage. Maturation culture was carried out in 60 x 15 mm Petri dishes . A cellulose membrane with 0.45 flm pore size (Millipore. Bedford. MA) was laid on a 35 x 35 x 5 mm polyester pad placed in the Petri dish. 6 mL maturation medium was added to the dish and absorbed into the polyester pad . This allowed the surface of membrane to be moist but not submerged with medium . Cells from week 3 ABA singulation culture were allowed to settle. The cells were washed with fresh maturation medium twice and then used for inoculum of maturation cultures. 1.2 mL of the cell suspension were transferred to the Petri dish and evenly laid on the surface of the membrane. The Petri dishes were sealed with parafilm and placed in the dark at 24 ±I qc.
Popu lation
l Il
r--G-IOb-u"-Iar--'I
Torpedo
2.2 1mage Analysis and Pattern Recognition The embryo images were obtained by a CCD camera and pre-processed according to the procedures described previously (Chi. Vits et aI. 1994; Vits . Chi et aI. 1994). In brief, the raw image was first binarized by defining a threshold value to discriminate embryos from the background. A thinning algorithm was used to extract embryo
Fig. l. Decison tree of the neural network based pattern recognition system (carrot).
486
embryos into three subgroups. Nodes #3 and #4 were used to separate the oblong+heart and The twins+c1uster groups, respectively . classification agreement between the two methods ranged from 73 to 88% for these four classes.
Table I . testing results of layer #1 of the neural network classifiers (carrot)
Normal Embryo
Abnormal Embryo
Neural Network Classification (%)
K·NN Method
Total Globular Torpedo Callus Others
~ ~
S2 Embryo
S3 Embryo
Globular
126
91.3
0.7
5.6
2.4
Torpedo
182
0.0
90.2
0.5
9.3
Callus
117
6 .0
0.0
93 .2
0.8
Others
403
3.7
6 .9
8.3
81.1
Fig. 2. Decison tree of the neural network based pattern recognition system (Douglas fir) Table 2 testing results of layer #2 of the neural network classifiers (carrot)
3. RESULTS AND DISCUSSION K·NN Method
Neural Network Classification (%)
3.1 Application of neural network for embryo classification.
Total Oblong +Heart
The classification results of each classifier for carrot embryos are shown in Tables I, 2 and 3 for the first. second and third layer classifiers, respectively. In each table, the second column contains the total number of embryos in a particular class as dassi fied previously using k·nearesl neighbor method . The other columns show the results of neural network The diagonal cells show the classification. agreement of the dassification between the two methods. At node #1 , the neural network dassifier separated the population into eight dasses, which were further grouped into four categories· globular. torpedo. callus and "Others" . "Others" contained oblong, heart. secondary, twins and cluster embryos . The agreement of globular, torpedo. callus and "Others" groups hetween the neural network classifier and K·NN method were 91.3 % . 90.2 %, 93.2% and X1.1 % . respectively. Group "Others" . consisting of fi ve embryo classes, was further
Secondary
Twins +Cluster
Oblong +Heart
151
92. 1
7.3
0.6
Secondary
108
19.4
57.4
232
Twins +Cluster
144
07
9.0
90.3
Table 3 Testing results of layer #3 of the neural net work c1assi fiers (carrot) K·NN Method
Total
Twins
Cluster
Twins
67
76. 1
23 .9
Cluster
77
11 .7
88.3
K·NN Method
separated into individual classes using node #2 (Figure I). The neural network classifier at this node was trained by another embryo population which contained only five embryo classes: oblong, heart, twins, cluster and secondary embryos. The embryos were separated into 3 subgroups : oblong+heart. twins+c1uster, and secondary embryos . Node #2 did not give a satisfactory separation of "Others" embryos from node # I into individual classes. Rather, it gave a satisfactory separation of those
Neural Network Classification ('To)
Neural Network Classification (%) Total
Oblong
Heart
Oblong
67
82.9
17 .1
Heart
77
26. 1
73.9
For Douglas fir. the population was first separated into two large groups - normal and abnormal embryos. A 3-layer feed forward neural network
487
classifier was used to accomplish this task. Ten sizerelated features and nine Fourier features (-5 to -2 and + I to +5) that contain the major morphological information of Douglas fir embryos were used as neural network classifier inputs.
Table 6: Comparison of classification of Douglas fir by a human operator and by the neural network classifiers. Human
Table 4 shows the testing results of the classifier at node #1 (Figure 2). The numbers in the second column are the human classification results and the last twO columns show the classification results of the neural network classifier. The classification accuracy for normal embryos was as high as 98 %. Further, the normal embryo group from node #1 was separated into three developmental stages corresponding to young, medium and mature embryos" This was achieved by a second neural network classifier at node #2 of the decision tree . The results are shown in Table 5. Among the 267 normal embryos classified by the operator 66, 136, and 65 fell into SI, S2, and S3 embryos, respectively. 54 out of 66 SI embryos were correctly classified by the neural network classifier, 11 classified as S2 embryos and I as S3 embryo. For S2 and S3 embryos, 118 out of 136 and 55 out of 65 were classified correctly giving a classification accuracy of 87% and 84% for S2 and S3 embryos, respectively. The accuracy of classifications was higher than 80% for all the morphological classes except abnormal embryos. The bottom two rows of Table 6 show the estimated number of embryos in each class and the accuracy of the neural network based pattern recognition system. The accuracy was from 89% for S3 embryos up to 97 % for S2 embryos. Overall , this pattern recognition system gave a good estimation of the embryos number in every class.
Normal Embryo Abnormal Embr;ro
Normal
Abnormal
Embr~o
Embr~o
271
267
4
32
10
22
81 82 83
Total 66 136
65
82
83
54
11 118 10
I
6
o
32
82
83
Abnormal
6 0 2
10 118 10 4
63
142
72
26
93 %
97%
89%
81%
2 2 0 22
12
55 4
ACKNOWLEDGEMENTS
Neural Network Classifier
81
Abnormal Total Accurac;r
81 55
Large variations in carrot embryo development were observed which is consistent with the population heterogeneity seen in batch embryo cultures. The rate at which each embryo progresses in the feature space was measured by a developmental vector. A major potential utility of single embryo tracking is to uncover a set of morphological features at an early embryo stage. that is associated with the eventuaI fate of those individuals. If such correlation could be found, then one can predict whether a particular embryo will become a mature. aborted. or abnormal embryo. even when it is in early stage. The analysis on developmental vectors revealed that embryos with a higher developmental rate during the carly stage of development had a higher chance of reaching the mature stage in a relatively short time . This single embryo tracking method is potentially a valuable tool in developing a correlation between the embryo's morphological features during the early stage of development and its final developmental fate .
Table 5: Testing at node #2 of the neural network classifiers (Douglas fu:1 Human Operator
68 138 65
The Fourier and size features span a multi-dimension feature space in which an embryo is represented by a particular point defined by its Fourier and size features. As the embryo develops, its Fourier and size features change accordingly and its position in the feature space also changes. This results in a series of points which give rise to a trajectory characterizing the developmental journey of this embryo. Connecting each point representing the individual embryos over time gives a trajectory which depicts the embryos' developmental "path" or history. By cultivating carrot embryos on Petri plates and following each individual embryo's development over time. we constructed such "developmental paths" for over 400 embryos.
Neural Network Classifier Total
81 82 83
3.2 Single embryo tracking
Table 4: Testing at node # 1 of the neural network classifiers (Douglas fir). Human Operator
Neural Network Classifiers
This work was supported in part by a grant from the National Science Foundation (BCS 9015817) We acknowledge Dr. Roger Timmis of Weyerhaeuser Corporation for kindly providing Douglas fir cell lines.
12
55
488
REFERENCES Chi, C. M., H. Vits, et al. (1994). Morphological kinetics and distribution in somatic embryo cultures. Biotechnol Bioeng. 44(3) : 368-378. Gupta, P. K., R. Timmis, et al. (1991). Field performance of micropropagated forestry species. In Vitro Cell. Dev. BioI., Plant 27P: 159-164. Huang, L. c., C. M. Chi, et al. (1993). Population and biomass kinetics in fed-batch cultures of Daucus carota L. somatic embryos. Biotechnol. Bioeng. 41(8): 811-818 . Taber. R. P .. C. Zhang. et al. (1997). Kinetics of Douglas fir (Pseudotsuga menziesii) somatic embryo development. Submitted. Timmis. R. (1995). Personal communication . Vits. H .• C.-M. Chi. et al. (1994) . Characterization of patterns in plant somatic embryo cultures : the morphology and development of embryos. AIChEJ . 40 : 1728-1740.
489