@2024 Afarand., IRAN

---

ISSN: 2251-8215    Sarem Journal of Reproductive Medicine 2017;1(3):97-104

Background
Anemia affects the shape, size, volume, color and number of red blood cells in a given volume. ... [1].‎
Previous Studies
‎... [2]. The usual methods of diagnosis of anemia, are clinical examinations of the patient and ‎observation of laboratory reports that contains important information including red blood cell count ‎and volume, hemoglobin concentration, etc. All of these calculations are performed by a cell counting ‎system in pathologic laboratories. Thousands of blood samples are taken daily in pathologic ‎laboratories around the world by microscopes and automatic counters. This is usually costly and time ‎consuming. Therefore, automatic detection of abnormal samples will be helpful in reducing time and ‎cost. One of the automated methods is laser scattering, which is used to determine the size of the red ‎blood cell, but has a high cost.‎
Aim(s)
The aim of this study was to measure the average volume of abnormal red blood cells using an ‎adaptive neuro-fuzzy inference system (ANFIS) by image processing.‎
Research type

Research society, place and time

Sampling method and number

Used devices&materials
The initial data used in the study was of a visual type, in the size of 640 × 480 and in the format of the ‎bitmap. The number of slides of normal people was 40 and the number of slides of abnormal ‎individuals was 20. A total of 30-50 images were taken from each slide, and a total of at least 1500 ‎normal images and 1000 unusual images were taken. Nikon's Microscope (Nicon Model NY100) and a ‎Panasonic camera were used to produce images in the laboratory. The camera was connected to the ‎computer from one side to the microscope and on the other side. The lens of the microscope was ‎adjusted on 40 and its magnification was adjusted to 100, and all images were prepared in the same ‎environmental and laboratory conditions. ... [3].‎ ‎ In the beginning, the color images turned to images with gray levels. The gray levels of the images were ‎in the range of 0.225. ... [4]. Threshold was used to convert images with gray levels to a binary image. ‎‎... [5].The red cells that stick to the edge of the image are considered unwanted objects and should be ‎removed from the calculations. Therefore, after the gray image became a binary image, these unwanted ‎objects were removed using the image retrieval method. This method uses a hybrid algorithm to ‎quickly recover images [6].‎ ‎ The objects tagging method was used to identify and specify the number of objects in the image. ... [7]. ‎To extract two characteristics of the diameter of the inner cavity of the red blood cell and the outer ‎diameter of the circle, the features of the smallest rectangle of the environment on the RBC were used. ‎Therefore, the average total length and width of this rectangle was considered as the outer diameter of ‎the red blood cell. The closest geometric shape to the human red blood cells is Torus, that this ‎geometric and mathematical simulation can be used to calculate the characteristics of the red blood ‎cell, including volume (Fig. 1) [8].‎ The Torus can be defined in forms of parameters in which the parameters u and v are in the range ( 0.2 π), R is ‎the distance from the center of the Turus to the center of the tube and r is the radius of the tube:‎ ‎1) y(u, υ)=(R + r cos υ ) sinus ‎2) z(u, υ)=r sin υ‎ This can be defined in Cartesian system:‎ ‎3) (R-√X2+Y2)2+Z2=r2‎ That it is easier to write:‎ ‎4) (x2+y2+z2+R2-r2)2=4R2(x2+y2)‎ A is the Torus area which is obtained from the following equation:‎ ‎5) A=4 π2 Rr=(2πr) (2πR)‎ And V, the volume enclosed by a torus is obtained from the following equation [8]:‎ ‎6) V=2π2 Rr2=(πr2) (2πR)‎ The volumes outside of the following range are considered unwanted and should be removed from the list before ‎the calculation is made.‎ ‎7) (a× median)≤Volume≤(b × median)‎ a and b are integers and a <1 and b>1. ANFIS was used to classify blood samples to normal and non-normal ‎groups. The ANFIS system is an adaptive and educable network that is similar in function to the fuzzy inference ‎system. For simplicity we assume that our fuzzy system has two inputs x and y and its output is z. Now if the ‎rules are as follows:‎ ‎8) Rule1: if x is A1 and y is B1 then ƒ1=Ƥ1x+q1y+r1‎ ‎9) Rule2: if x is A2 and y is B2 then ƒ2=Ƥ2x+q2y+r2‎ And if for the non-fuzzy makers, we use the non-fuzzy maker of the average of the centers, the output will be as ‎follows:‎ ‎10) ) f=(w_1 f_1+w_2 f_2)/(w_1+w_2 ) w ̅_1 f_1+w ̅_2 f_2 stw ̅_1=w_1/(w_(1+) w_2 ),w ̅_2=‎w_2/(w_1+w_2 )‎ Layer Output: Layer 1: In this layer, inputs pass through membership functions:‎ ‎11) O1,i=μAi(x)ƒor i=1,2‎ ‎12) O1,i=μBi(x)ƒor i=3,4‎ The membership functions can fit any parametric function, which in most cases is chosen as Gaussian functions. ‎Such as a generic bell function that is a set of parameters. Parameters of this layer are known as early parameters.‎ ‎13) μA(x)=1/(1+|(x-c_i)/a_i |^(2b_i ) )‎ Layer 2: The output of this layer is the multiplication of the input signals, which is actually equivalent to the "if" ‎part of the rules.‎ ‎14) O2,i=wi=μAi (x) μBi (y),i=1,2‎ Layer 3: The output of this layer is normalization of the previous layer ‎15) O_(3,i)=w ̅_i=w_i/(w_1+w_2 ),i=1,2‎ Layer 4: Non-fuzzy average center maker ‎16) O4,i=w̅ i fi=w̅ i (pi x+qi y+ri Layer 5:‎ ‎17) O_(5,i)=∑_i▒w ̅_i f_i=(∑_i▒〖w_i f_i 〗)/(∑_i▒w_i )‎ Now a network has been generated which is equivalent to the TSK fuzzy inference system (Fig. 2) [9].‎ The final output of the fuzzy system is shown in Fig. 3 [9].‎ Therefore, in this network, the Gaussian multiplication and Gaussian fuzzy, and the center average non-‎fuzzy maker were used. To train the adaptive neuro-fuzzy network parameters, a combination learning ‎algorithm was used [9]. The REC (Receiver Operating Characteristic) curve was used to evaluate ‎performance (sensitivity and test characteristics) in a system. The sensitivity of a test is its ability to ‎detect a disease within a patient's population and, in fact, is the correct diagnosis of patients by the ‎test.‎ ‎21) TPF=TP/(TP+FN)‎ Number of patients= (TP + FN)‎ The extent of the characteristics of a test is its ability to detect the absence of a disease in a disease-free ‎population; in fact, it can be said that the rate of the number of people who are not patient and tested by ‎the test have been correctly recognized.‎ ‎22) TNF=TN/(TN+FP)‎ Number of non-patient persons = (TN + FP)‎ The accuracy of the test is the correct rate of the results of a test. In fact, the number of people who are ‎correctly diagnosed () is divided into the population of all individuals.‎ ‎23) (TP+TN)/(TP+FP+TN+FN)‎ The area under the ROC curve shows the system's efficiency. The largest area is one (Fig. 4).‎ When the efficiency of a test increases, the ROC curve moves to the upper left corner of the graph, the ‎area under the curve also increases and approaches 1 [10].
Findings by Text
Initially, the color image turned into an image with gray levels. The red cells that were attached to the ‎edge of the image were removed using the image processing method (Fig. 5).‎ The location coordinates of each object were used to extract the following image with gray levels from the ‎original image. Therefore, each gray underneath image contained only one cell. Then, each beneath of the ‎image with gray levels was converted to binary image beneath using the otsu algorithm with the optimal ‎threshold (Fig. 6).‎The size of the inner and outer diameters of the red cells was calculated. To remove unwanted ‎volumes, volumes that were in the middle of 50% of mode were selected as acceptable volumes. The ‎cross-sectional evaluation method was used to select and reverse the training data and test data. ‎Therefore, the neuro-fuzzy network was tested once for all normal and abnormal data in the system. ‎The training data included 21513 red cells (13835 normal data and 7678 abnormal data).‎ The outer radius and radius of the red cell cavity were used as two input variables in the neuro-fuzzy ‎network. For each entry, the Gaussian membership function is assigned, so the rule base consists of 36 ‎bases. After 100 rounds of training, the mean squares of training error and test were 0.005 and 0.307 ‎respectively. The form of the membership functions of the two inputs of the network after the ANFIS ‎training and the procedure of the inputs and outputs of the network respectively are shown in Figures 7 ‎to 10. Desired output and actual output of the ANFIS network based on the test data belonging to one ‎category (each category including 4 normal people and 2 non-normal people) is shown in Fig. 11. The ‎accuracy of the system and the area under the ROC curve were 96.6% and 99.5%, respectively.‎ The numerical results from the implementation of the neuro-fuzzy network are shown in Table 1.‎
Main comparison to the similar studies
‎… [11]. In 2009, Apostolopoulos et al. presented a method for estimating the size of human red blood ‎cells using images containing scattered light. The image data recovery process involved normalizing ‎the image, applying a two-dimensional discrete cosine transform to the image, and applying a wavelet ‎transform to the image. An RBF-NN (RBF-Neural Network) neural network, which estimated the ‎geometric properties of red blood cells, was investigated. The proposed method has been used when ‎three important RBC geometric features have been investigated using a database of 1575 simulated ‎image with a threshold method [12]. ‎
Suggestions

Limitations

Conclusions
The proposed method provides diagnostic capability using a drop of blood and is well suited to ‎pathological images. The designed automatic system can be a convenient and affordable alternative to ‎commonly used laboratory procedures. In addition, the proposed method can be a base for calculation ‎of other parameters of blood test or CBC such as CBC, HCT, MCH, MCHC and the number of red blood ‎cells, etc. ‎
Acknowledgments
This article is the result of a joint research carried out between the Cellular-Molecular Research Center, ‎the Stem Cells of the Sarem Medical Faculty, Center of Research, Sciences and Technology in Medicine ‎of Tehran University, Amir Kabir University of Technology, and the Azad University of Science and ‎Research Branch. All the relevant colleagues who helped us in conducting this study are appreciated. ‎
Conflict of interest

Ethical Permissions

Funding Sources

TABLES and CHARTS

Show attach file