Incorporation of contextual information in object identification

[0002] Priority to provisional U.S. patent application Ser. No. 60/063,690, filed Oct. 28, 1997 is claimed.

[0003] The invention relates to techniques of identifying and classifying objects. More specifically, the present system describes using contextual information to improve statistical accuracy.

[0004] A common assumption made in the field of machine learning is that the examples are drawn independently from some joint input-output distribution.

[0005] Classical statistics begins with an assumption of a population distribution. Tiat population distribution describes how some attribute varies over a set of entities. The goal is to obtain information about the population distribution by sampling or by some other real world technique. It is often impractical to evaluate the parameters exactly. Therefore, one attempts, using the techniques of statistics, to obtain as much information as one can based on experimental evidence. Then all of the parameters can be viewed a coordinates of a single vector often called x . The value of x is fixed, but there is uncertainty about its value.

[0006] The population often has a distribution. For example, if the population has a Gaussian distribution, the parameters may have a mean value μ and a standard deviation Σ.

[0007] The Bayesian approach models the population by considering elements of the population to be randomly selected from a class of populations. Each class of the population is characterized by parameters that are under investigation. Bayesian analysis considers two parameters x and y. Bayesian analysis is largely an effort at determining the probability for the occurrence of the event y=β conditioned by the fact that x=α has occurred. During N trials the event x=α occurs n_x(α) times. The joint effect x=α and y=β occurs n_xy(α, β) times.

[0008] The usual Bayesian approach assumes the distribution for x by looking at the trials for x. Hence, Bayesian analysis considers the trials in which x=α. Once the random sampling processes are obtained, the notion of the random sampling process is extended by assuming that the class x_i: 1 is less than or equal to I; is conditionally independent. Each of the x_ihave the same conditional distribution. This is because each of the x_iare defined through the same random parameter x. While the events x and y are independent, both have the same conditional probabilities. Hence, the conditional probability that y=β given that x=α is independent of α. This means that $P_{x|y}\left(\alpha |\beta \right)=\frac{P_{x|y}\left(\beta |\alpha \right)P_x\left(\alpha \right)}{P_y\left(\beta \right)}$

[0009] where the left part is the conditional probability that y=β given that x=α, the right part is the conditional probability that x=α given that y=β, p_x(α) is the probability of α and p_x(β) is the probability of β.

[0010] Bayesian analysis is often used in estimator design. For instance, the conditional probability of a received message may be known and the conditional probability of the transmitted message is to be obtained. Bayesian analysis requires that the examples are drawn independently from the joint input/output distribution, as described above.

[0011] There are cases, however, where this no context assumption is not valid. One application where the independence assumption does not hold is the identification of white blood cell images. Abnormal cells are much more likely to appear in bunches than in isolation. Specifically, in a sample of several hundred cells, it is more likely to find either no abnormal cells or many abnormal cells than it is to find just a few. More generally, certain random processes are subject to “aliasing” whereby they occur in clumps.

[0012] The present specification describes pattern classification in situations where an independence assumption is not necessarily correct. In such cases, the probability of identity of an object may be dependent on the identities of the accompanying objects. Those accompanying objects which provides the contextual information.

[0013] The techniques take into consideration the joint distribution of all the classes, and uses the joint distribution to adjust the object-by-object classification.

[0014] This method can be used in a number of settings. For example, the classification of white blood cells is improved by the use of contextual information.

[0015] These and other aspects of the invention will now be described with reference to the attached drawings, in which:

[0016] FIG. 1 shows exemplary views of different kinds of white blood cells;

[0017] FIG. 2 shows a flowchart of no-context classification;

[0018] FIG. 3 shows a device for carrying out this operation; and

[0019] FIG. 4 shows the post processing operation to include contextual information.

[0020] The objects are described herein by a vector that defines their probabilistic characteristics taken as having feature vectors.

[0021] The feature vector is defined as x_i. The feature vector is be classified into one of a number of classes. The classification of a given feature vector x_iis defined as c_i=c(x_i), I . . . 1 to N, where N is the total number of objects and hence the total number of feature vectors. The classification c_iis a member of the set {1, . . . , D}, where D is the total number of classes.

[0022] For example, if one considers ten samples of blood, possible feature vectors could include the different types of white blood cells.

[0023] Leukocyte analysis is an important routine laboratory examinations. Leukocyte classification aids in clinical diagnosis because the relative percentage composition of the blood leukocytes changes in various physiological and pathological conditions. An estimate of the percentage of each class present in a blood sample conveys information which is pertinent to the hematological diagnosis.

[0024] Typical commercial differential white blood cell (“WBC”) counting systems are designed to identify five major mature cell types. Blood samples, however, may also contain other types of cells, i.e., immature cells. These cells occur infrequently in normal specimens. Most commercial systems will simply indicate the presence of these cells because such systems cannot be individually identified. These cell types that relate to the production rate and maturation of new cells. The present inventors, however, believe that these are important indicators of hematological disorders.

[0025] The preferred system differentiates up to fourteen WBC types. This includes the five major mature types: segmented neutrophils, lymphocytes, monocytes, eosinophils, and basophils; and the immature types: bands (unsegmented neutrophils), metamyelocytes, myelocytes, promyelocytes, blasts, and variant lymphocytes; as well as nucleated red blood cells and artifacts. Differential counts are made based on the cell classifications, which further leads to diagnosis or prognosis. The fourteen different cell types to be identified are shown in FIG. 1.

[0026] According to Bayes rule, where p(c|x) is the probability of c if x is given, p(c) is the probability of c and p(x) is the probability of x. $p\left(c|x\right)=\frac{p\left(x|c\right)p\left(c\right)}{p\left(x\right)}$

[0027] That is, the probability of c given x, is the probability of x given c, times the probability of c divided by the probability of x.

[0028] This is the no-context rule, correct only when the entire class has a random and flat probabilistic value. In white blood cell analysis, as described above, the probability of any sample being any c is dependent on the rest of the neighboring population. In context, each of the classes of this system can have a different probabilistic value. Therefore, the “with context” a posteriori probability of the class label of all the objects assumes value C₁, C₂, . . . , C_N, given all the feature vectors, is $\begin{array}{cc}p\left(c_1,c_2,\dots ,c_N|x_1,x_2,\dots ,x_N\right)=\frac{p\left(x_1,x_2,\dots ,x_N|c_1,c_2,\dots ,c_N\right)p\left(c_1,c_2,\dots ,c_N\right)}{p\left(x_1,x_2,\dots ,x_N\right)}& \left(1\right)\end{array}$

[0029] The feature distribution of a given class is independent of the feature distributions of other classes, i.e.,

[0030] p(x₁x₂, . . . ,x_N|c₁,c₂, . . . ,c_N)=p(x₁|c₁) . . . p(x_N|c_N)

[0031] This Equation (1) can be rewritten as $\begin{array}{cc}p\left(c_1,c_2,\dots ,c_N|x_1,x_2,\dots ,x_N\right)=\frac{p\left(x_1|c_1\right)\dots p\left(x_N|c_N\right)p\left(c_1,c_2,\dots ,c_N\right)}{p\left(x_1,x_2,\dots ,x_N\right)}=\frac{p\left(c_1|x_1\right)\dots p\left(c_N|x_N\right)p\left(x_1\right)\dots p\left(x_N\right)p\left(c_1,c_2,\dots ,c_N\right)}{p\left(c_1\right)\dots p\left(c_N\right)p\left(x_1,x_2,\dots ,x_N\right)}& \left(2\right)\end{array}$

[0032] Where p(c_i|x_i) is the “no context” object-by-object Bayesian a posteriori probability, p(c_i) is the a priori probability of the classes, p (x_i) is the a priori probability of the features. The features (x₁, x₂, . . . , x_N) are given. Hence their probabilities p(x₁, x₂, . . . , x_N) and p (x_i) are constant. Therefore $p\left(c_1,\dots ,c_N|X_i\dots X_n\right)\propto p\left(c_1|x_1\right)\dots p\left(c_N|x_N\right)\frac{p\left(c_1,\dots ,c_N\right)}{p\left(c_1\right)\dots p\left(c_N\right)}$

[0033] where

[0034] The description made herein is for the application of object $\begin{array}{cc}=p\left(c_1|x_1\right)\dots p\left(c_N|x_N\right)p\left(c_1,c_2,\dots ,c_N\right)p\left(c_1,c_2,\dots ,c_N\right)\equiv \frac{p\left(c_1,c_2,\dots ,c_N\right)}{p\left(c_1\right)\dots p\left(c_N\right)}& \left(3\right)\end{array}$

[0035] classification. Object classification determines the count in each class, rather than the particular ordering or numbering of the objects. As a result, P(c₁, c₂, . . . , c_N) is only a function of the count in each class.

[0036] Let N_dbe the count in class d, and u_dis the percentage of objects in the class, =N_d/N, for d=1 . . . ,D, then: $\begin{array}{cc}p\left(c_1,c_2,\dots ,c_N\right)=\frac{p\left(c_1,c_2,\dots ,c_N\right)}{p\left(c_1\right)\dots p\left(c_N\right)}=\frac{N!p\left(v_1,v_2,\dots ,v_D\right)}{N_1!\dots N_D!P_1^{{\mathrm{Nv}}_1}\dots P_D^{N_{\mathrm{VD}}}}& \left(4\right)\end{array}$

[0037] where P_dis the prior distribution of class d, for d=1, . . . D. $\begin{array}{c}\sum _{D=1}^D\mathrm{Nd}=N\mathrm{and}\sum _{D=1}^D\mathrm{vd}=1.\\ =p\left(v_1,\dots ,v_D\right)\end{array}$

[0038] =p(v₁, . . . ,v_D)

[0039] (c₁,c₂, . . . ,c_N)

[0040] The decision would then become $\begin{array}{c}\left(c_1,c_2,\dots ,c_n\right)=\arg \max p\left(c_1,c_2\dots ,c_N|x_1,x_2,\dots x_N\right)\\ \left(c_1,c_2,\dots ,c_n\right)\end{array}$

[0041] Implementing this decision rule requires computing and comparing D^Nterms to solve Equation 5. In our case of white blood cell recognition, D=14 and N is typically around 600, requiring 600¹⁴≈8×10³⁸terms—an unmanageable number of calculations.

[0042] The technique disclosed uses a simplification to avoid the massive computation. Instead of considering all the classes all at once, one class is considered at a time. Take the class d for example.

[0043] Defining p^d_ito denote the probability that d is the classification for group x_i;=P(c_i=d|x_i), I=1, . . . ,N. We order the P^d₁, P^d₂, . . . , P^d_N's in a descending manner dependent on I, such that

[0044] p₁^d>=p₂^d>= . . . >=p_N^d  (6)

[0045] Then, the probability that there are n cells belonging to class d is $P_d\left(n\right)=p_1^d\dots p_n^d·\left(1-p_{n+1}^d\right)\dots \left(1-p_N^d\right)p\left(v_1,\dots ,v_d=\frac{n}{N},\dots ,v_D\right)$

[0046] The P_d(n)'s can be computed recursively.

[0047] First, we calculate P_d(0): $P_d\left(0\right)=\left(1-p_1^d\right)\dots \left(1-p_N^d\right)p\left(v_1,\dots ,v_d=\frac{0}{N},\dots ,v_D\right)$

[0048] Then, the other values are calculated from Pd(0) $P_d\left(n+1\right)=P_d\left(n\right)\frac{p_{n+1}^{d}p\left(v_1,\dots ,v_d=\frac{n+1}{N},\dots ,v_D\right)}{1-p_{n+1}^d)p\left(v_1,\dots ,v_d=\frac{n}{N},\dots ,v_D\right)}$

[0049] This way all the P_d(n)'s computed from only N terms.

[0050] The value n=n* that maximizes P_d(n), n=1, . . . , N is selected. Then all the objects that rank highest—the top n* in Equation 6—are classified into class d. This can be carried out in a similar way for all the classes.

[0051] Importantly, $p\left(v_1,\dots ,v_D\right)=\frac{N!p\left(v_1,v_2,\dots ,v_D\right)}{N_1!\dots N_D!P_1^{{\mathrm{Nv}}_1}\dots P_D^{{\mathrm{Nv}}_D}}$

[0052] is simply a count of overall probability of any particular white blood cell. This count can be estimated from a database of known cells.

[0053] Application to White Blood Cell Recognition

[0054] The data was provided by IRIS, Inc. (International Remote Imaging Systems, Inc., Chatsworth, Calif.). Blood specimens were collected from local patients at Harbor UCLA Medical Center, then dyed with Basic Orange 21 metachromatic dye supravital stain. The specimens were then passed through a flow microscopic imaging and image processing instrument, where the blood cell images were captured. Each image has a single cell with full color. There are typically 600 images from each specimen, so N−600.The cell recognition system categorizes the cells based on the images.

[0055] The size of cell images are automatically tailored according to the size of the cell in the images. Images containing larger cells had bigger sizes than those with small cells. The range varied from 20×20 to 40×40 pixels. The average size is about 25×25 pixels.

[0056] The image processing operation is described with reference to the flowchart of FIG. 2, in which steps 200 and 202 represent the cell-by-cell classification of each cell. Each image containing a single cell is segmented in order to emphasize the interior of the cell relative to the image. At step 202 the features are extracted. These features which are extracted include, for example, size, shape, color and texture. The color image is decomposed into its three primary colors, red, green and blue. The portions of red, green and blue set the overall color that is used for identification. Since the cells fall into certain categories, one of the processing features used determines red and blue information. The first part of the operation determines a red-blue distribution. This is the pixel-by-pixel value of log(red)—log(blue) which forms the distribution for the pixels in the cell interior. The red distribution from pixels in the cell interior can also be obtained. The preferred featured information, including the size, shape, color and texture, are described in more detail in Table 1.

[0057] The features are analyzed using a neural network shown in FIG. 3. A nonlinear feed-forward neural network 301 has 20 inputs, 15 hidden units with sigmoid transfer functions, and 14 sigmoid output units. This architecture is chosen through cross-validation. A cross-entropy error function is used in order to give the output a probability interpretation.

[0058] Denoting the input feature vector as x, the network outputs a D-dimensional vector, where D=14 in this embodiment p={p(d|x)}, d=1, . . . , D, where p(d|x) is the probability that a cell belongs to class d, given the presence of feature x. The decision made at this stage is

[0059] d(x)=argmaxp(d|x)

[0060] Once the cell-by-cell classification is made, the present system uses the post processor 305 to refine this cell-by-cell classification using the contextual information.

[0061] Each of these features identified in step 202 becomes a feature number which is used to help identify the cell and where it belongs relative to the other cells. Comparisons and associations are also used to compare the cell under interest with other cells in the same specimen. According to this embodiment, the cell is more likely to assume a certain identity if the company it keeps supports that identity.

[0062] Again, the “no-context” cell-by-cell decision is only based on the features presented by a cell, without looking at any other cells. However, a cell is more likely to assume a certain identity if the surrounding cells support that identity.

[0063] For instance, the difference between lymphocyte and blast can be very subtle sometimes, especially when the cell is large. A large unusual mononuclear cell with the characteristics of both blast and lymphocyte is more likely to be a blast if it is surrounded by or accompanied by other abnormal cells or abnormal distribution of the cells.

[0064] This type of analysis fits within the framework presented above for incorporating contextual information. Hence, the image 300 is also coupled to post processor 305. The post-processing draw on the mathematical framework above. The overall technique proceeds according to the basic probability flowchart of FIG. 4.

[0065] At step 400, the algorithm estimates P(v, , . . . v_D) from the database for each of d=1 . . . D. Since N_dis the percentage of objects in class D, the numbers can in fact be estimated.

[0066] At step 402, the object-by-object no context, a posteriori probability p(c_i|x_i) is computed for all objects I=1 . . . N, C∈1 . . . D. This uses known statistical techniques described in equation (2). Now all terms for equation 7 have been calculated.

[0067] At step 404, the probability Pd(n) that a cell belongs to class D is calculated using equation (7) using the terms calculated in steps 400 and 402. The maximum value n_d* for all D's is also determined. The cells are re-labeled appropriately at step 406. This corresponds to the equation 6.

[0068] Step 406 is continued until the re-labeling process converges, shown as step 408.

[0069] The technique described above for combining contextual information is used as the post-processing step using post processor 405 for the cell-by-cell decision process. The final output information then is post-processed using techniques described above.

[0070] The techniques described herein have been intensively tested at IRIS, Inc. on the specimens obtained at Harbor UCLA medical center. The performances with or without using contextual information on blood samples from 200 specimens (13,200 cells) were tested. In 50% of the cases, a false alarm would have occurred had context not been used. Most cells are correctly classified, but a few are incorrectly labeled as immature cells, which raises a flag for the doctors. Change of the classification of the specimen to abnormal requires the expert intervention before the false alarm is eliminated, and it may cause unnecessary worry. When context is applied, the false alarms for most of the specimens were eliminated.

[0071] The technique described can be generalized to a number of domains where contextual information plays an essential role, such as speech recognition, character recognition, and other medical diagnosis regimes. In speech recognition and character recognition, the identity of a word depends on the identities of neighboring words. Context plays a key role, much as it does in the recognition of blood cells.

[0072] Although only a few embodiments have been described in detail above, other embodiments are contemplated by the inventor and are intended to be encompassed within the following claims. In addition, other modifications are contemplated and are also intended to be covered.