You are on page 1of 4

Estimation Of Hemoglobin Using A I Technique

Suhas B. Dhoke1, Anil karwankar2, Varsha Ratnaparkhe3


Electronics & Telecommunication Engineering Department,
Government College of Engineering, Aurangabad
1
suhasdhoke84@gmail.com, 2karwankar@yahoo.com,3patwadkar.varsha@gmail.com

Abstract—Anemia is a condition in which the hemoglobin (Hb) available. To overcome complication related to anemia. We
content becomes less than that of the normal value. In this project, need to develop a system which will able to identify
hemoglobin value is estimated using ANN (Artificial Neural hemoglobin level in earlier stage of anemia very easily. So that
Network). Database of blood sample images and their actual Hb it will help to take care of anemia patient. There are various
values is collected from a local laboratory. Red, green and blue methods available for the estimation of Hb level in blood. These
normalized values of images’ samples are fed to the ANN as input. methods are documented by the World Health Organization
Cyanemethemoglobin method based calculated values of Hb (WHO) [5]. Following is list of the methods for estimating Hb
obtained from the laboratory are given as output. Comparing the value.
outputs of ANN model results with actual Hb values, accuracy of
the network is calculated. This paper covers comparison of TABLE 2: DIFFERENT METHODS OF ESTIMATIN HB LEVEL
performance of different types of Neural Networks for carrying
out the stipulated task. It is observed that there is a strong relation Sr. No. Methods for estimating Hb Accuracy
between red, green and blue color components of the image with 1. Filter paper method Less
the hemoglobin content of the blood.
2. Pallor test method Less
Keywords—Anemia, hemoglobin, cyanemethemoglobin, ANN
3. Lovibond-type comparator Moderate
I. INTRODUCTION
Hemoglobin is a protein molecule in red blood cells that
4. Copper sulphate method Moderate
contains iron as the main element. Hemoglobin is made up of 5. Sahli’s method Moderate
globulin chains that is interconnection of protein molecules.
Heme is one of the important component of globulin chain, 6. HemoCue method High
which contain iron atom. Its help to transfer O2 and CO2 in our
body through the blood [1]. Red color of blood is because of the 7. Cyanemethemoglobin Very high
iron contained in hemoglobin molecules. It means if there is
more concentration of hemoglobin in blood then there would be
more red component within the image samples of blood. Amongst these methods, Cyanemethemoglobin method is
considered the most accurate and is accepted worldwide but it
Anemia is found to be a common health problem in women, need to handle the hazardous elements like cyanide, some
teenage girls and children. In India, as per the figures of unique equipment with a skilled technician. That’s why it’s not
National Family Health Survey - 4, 52.5% women in the age considers as usual method for identifying hemoglobin level in
group of 15 to 49 years and 25.5 % men are suffering from blood [6].. The method based on an ANN can give us possible
anemia [2]. Normal level of Hb in the blood is as given below accurate output for the given blood sample.
in Table 1 [3].
TABLE 1: RANGE OF HEMOGLOBIN IN BLOOD
II. ARTIFICIAL NEURAL NETWORK
ANN is a structural model. There is interconnected web of
Sr. No. Category Range of Hb in blood nodes called as neurons and line to join them which contain its
own weight. Main function is to receive the input, perform
1 Women 12 -16 gm/100 ml progressively complex calculation and then use output to solve
2 Men 14 -18 gm/100 ml the problem [7]. Due to this self-learning ability of neural
networks the optimum weight assignment leads to satisfactory
3 Newborn 14 -20 gm/100 ml results in pattern classification problems.
General procedure of implementation of ANN and flow of
ANN is shown in the figure 1 [8].
Anemia is a condition in which the Hb content is less than
the normal value. If there is deficiency of iron, folic acid or
vitamin then that would be result it into the decrease INITIALISE THE
NETWORK
hemoglobin level in the blood. The problem of anemia occurs
specifically during pregnancy, menstruation and sport
activities. [4]. Most of people lives in rural area in developing
country and proper medical facilities are not available there, NETWORK TRAINIG
like analysis of blood samples. That’s why there is need to
develop system which will able to measure the hemoglobin
content in the blood. That system have to be more accurate and
easy to operate. We can take preventing measure before getting OUTPUT OF THE NETWORK
COMPARE WITH DESIRE
worst anemia condition, accordingly to Hb level in blood OUTPUT
patients would need to intake the nutrition. This could be
possible if easily operated Hemoglobin estimation system are Fig.1 Flow of ANN

978-1-5090-4264-7/17/$31.00 ©2017 IEEE


54
III. IMPLEMENTATION & PERFORMANCE Hb obtained after training neural network for respective
samples. And green line shows that difference of output value
A. Normalization of image of ANN after the training and actual value of output for
There could be chance that the images of blood samples are respective samples. After the training, that difference ideally
obtained at different lighting conditions. At high intensity light should be zeros.
there would be rise in the red, green, blue components’ values
of image while at low light condition, red, green, blue
components’ values may get reduced. To avoid such different
lighting conditions, the image is needed to be normalized.
The image is normalized as follows:


= − +

Max: maximum pixel intensity


Min: minimum pixel intensity
IN: normalized pixel intensity
I: original pixel intensity
newMax: new maximum value Fig. 3. Neural network training regression plot
newMin: new minimum value
Figure 3 is neural network training regression plot. It shows
B. Backpropagation network the target values which are given to the network as output, on
Backpropagation algorithm repeats two phases i.e. propagation x- axis and the actual output of neural network after training.
and weights updates. When input is given to network, it
propagates through each of its layers until it reaches to its
output. This output is compared with desired output and for
each neuron in output layer, error value is calculated, then error
value propagates backwards. This process is continued
iteratively.

Iron contributes to the Hb level of blood and makes it red. There


is a strong relation between the color components and the
hemoglobin. In this paper, Hb value is estimated by using the
color components of images samples. The average of red
component of colored image is computed and considered as
feature. These are calculated for the dataset available.
Features are given to the ANN as input. ANN is designed with
a three layer back propagation network. The laboratory given
Hb values are given as output. Obtained results are indicated
diagrammatically as below:
Fig. 4. Neural network training performance

Figure 4 shows the variation of value of mean squared error on


iterations during the training. After 600 iterations, curve seems
to be a nearly straight line. From this curve it is clear that if the
number of iterations are increased further, it will not make the
big change in mean square error. So instead of increasing the
iterations, number of neurons in each layer needs to be
increased.
By doing so, following results are obtained:
Figure 5 indicates Hb versus number of samples. It just
indicates the actual Hb values, output values of ANN and the
difference between these. Time required for training is a little
bit higher due to increase in the number of neurons but more
accurate results than previous one were obtained. Figure 6
depicts the target Hb values and the output values and Figure 7
indicates reduction of mean square error with respect to epochs.
Fig. 2. Hb value vs number of samples

In figure 2. The Blue line shows how the value of Hb varies


with respect to each sample, while red line shows the values of

55
Figure 10 indicates the graph of actual Hb values obtained from
ANN and laboratory values of Hb.

Fig. 5. Hb value vs number of samples

Fig. 8. Hb value vs normalized mean value of red component

Fig. 6. Neural network training regression plot

Fig 9. Neural network training performance

Fig 7. Neural network training performance

From the curve in figure 7, it can be concluded that there may


be a chance to improve the network parameter by increasing the
number of iterations. Fig. 10. Neural network training regression plot

C. Leverberg-Marquaratt function netwrok In figure 8, as green line is going to be straighter it shows that
there is more accurate result, and from figure 9 curve is going
The Levenberg-Marquardt algorithm is simple, efficient and to be decreasing. So by increasing the number of iterations
robust for approximating a nonlinear function. there could be chance of improving the results.

To increase the computational speed, Leverberg-Marquaratt D. Radial basis function network


function network, as the back propagation (gradient) is known In Radial basis function network, there is requirement of more
to be slow. Less number of neurons are used now. Following number of neuron than back-propagation network and one of
results were obtained. Figure 8 exhibits actual and obtained the important advantage is that it can be designed in very less
outputs variations and difference of the two. Figure 9 depicts amount of time. If there is more number of non-linear training
the reduction of mean square error with number of samples. element are available then Radial basis function network would

56
be the best choice. Specialty of radial basis network is that it IV. CONCLUSION
can fit any function with any given accuracy with the help of After training neural network with above methods, these neural
correct weight and bias values for each layer. Radial basis network are tested by the different blood image samples means
network has two layers; one is hidden layer of radial basis other than the training samples. Different values of accuracy are
neurons while the other is output layer of linear neurons. obtained which are enlisted in table 2.
Input is given to radial basis function network. Number of TABLE 3: ACCURACY
neurons in the layers is decided by its optimization algorithm
which is found to be as follows. Figure 11 shows Radial basis Sr. Methods Accuracy
network. Figure 12 shows number of samples and No. in testing
corresponding Hb values. Red and blue line indicates the %
laboratory values and actual values of Hb. Green line show the 1 Back propagation network 71 %
difference of these two.
2 76 %
Increase number of neurons
3 Leverberg Marquaratt function 86 %
network
4 Radial basis function network 91 %

It is found that configuration with radial basis function network


Fig. 11. Radial basis network has the best results. In case of back propagation network, it does
not offer good accuracy and it takes long time for training the
network. L-M function network seems to be more efficient in
non-linear approximation function. Time require for training
the L-M function is less compare to back propagation. It is
observed that increasing number of neurons and number of
layers minimized the error, while time required for training
increased. One more thing observed that instead of the giving
the direct value of Hb to neural network training, if we make
class of Hb value then given to network as output for training
then there is more chance of rise in the accuracy because we
approximating the output values in different classes.
REFERENCES
[1] Non-Invasive Estimation of Hemoglobin in Blood Using Color Analysis
First International Conference on Industrial and Information Systems,
ICIIS 2006, 8 - 11 August 2006, Sri Lanka
Fig. 12. Hb value vs number of different samples [2] F. Halsall, Multimedia Communications. Delhi, India: Pearson Education,
2001.
Figure 13 depicts the comparative data of laboratory values [3] B. De Benoist, E. McLean, I. Egli, and M. Cogswell, “Who global
and actual ANN output values. database on anaemia,” Geneva: WHO, pp. 1993–2005, 2008.
[4] M.Malhotra, “Severe anemia linked to poor outcomes for pregnantwomen
and their babies,” Int. J. Gynecol. Obstet., vol. 79, pp. 93–100, 2002..
[5] H. Ranganathan and N. Gunasekaran, “Artificial Neural Network
Approach in Estimation of Hemoglobin in Human Blood Using Color
Analysis”, IEEE Transactions On Information Technology In
Biomedicine, vol. 10, no. 4, October 2006.
[6] S. Ghosh and M. Mohan, “Screening for anemia,” Lancet, vol. 1, p. 823,
1978..
[7] Li, J., A. N. Michel, and W. Porod, “Analysis and synthesis of a class of
neural networks: linear systems operating on a closed hypercube,” IEEE
Transactions on Circuits and Systems, vol. 36, no. 11, November 1989,
[8] NEURAL NETWORKS: Basics using MATLAB Neural Network
Toolbox By Heikki N. Koivo on February 1, 2008

Fig. 13. Neural network training regression plot

In figure 13, almost all the points are on the blue line, which
indicates that most accurate results are obtained with the radial
basis function network.

57

You might also like