Machine learning approach of automatic identification and counting of blood cells

Mohammad Mahmudul Alam, Mohammad Tariqul Islam ✉

Department of EEE, Bangladesh University of Engineering and Technology, Dhaka, Bangladesh
✉ E-mail: tariqul@[Link]

Published in Healthcare Technology Letters; Received on 20th October 2018; Accepted on 3rd May 2019

A complete blood cell count is an important test in medical diagnosis to evaluate overall health condition. Traditionally blood cells are counted
manually using haemocytometer along with other laboratory equipment’s and chemical compounds, which is a time-consuming and tedious
task. In this work, the authors present a machine learning approach for automatic identification and counting of three types of blood cells using
‘you only look once’ (YOLO) object detection and classification algorithm. YOLO framework has been trained with a modified configuration
BCCD Dataset of blood smear images to automatically identify and count red blood cells, white blood cells, and platelets. Moreover, this study
with other convolutional neural network architectures considering architecture complexity, reported accuracy, and running time with this
framework and compare the accuracy of the models for blood cells detection. They also tested the trained model on smear images from a
different dataset and found that the learned models are generalised. Overall the computer-aided system of detection and counting enables
us to count blood cells from smear images in less than a second, which is useful for practical applications.

1. Introduction: A complete blood cell (CBC) count is an Also, the trained model has been tested with images from another
important test often requested by medical professionals to dataset to observe the generalisation of the method. Fig. 1 shows
evaluate health condition [1, 2]. The main three types of cells that the proposed deep learning based blood cell identification and
constitute blood are red blood cells (RBCs), white blood cells counting system.
(WBCs), and platelets. RBCs also known as erythrocytes
are the most common type of blood cell, which consists of 2. Related works: In general, there are generally two different
40–45% of blood cells [American Society of Haematology: approaches in the automated counting process of blood cells.
[Link] Platelets also known They are the image processing approach [1, 3, 10–12] and the
as thrombocytes are also in huge number in blood. WBCs also machine learning approach [2, 4, 13–15].
known as leukocytes, are just 1% of total blood cells. RBCs carry Acharya and Kumar [10] proposed an image processing
oxygen to our body tissues and the amount of oxygen tissues technique for RBCs count. It processed the blood smear image to
receives is affected by the number of RBCs. WBCs fight against count RBCs along with the identification of normal and abnormal
infections and platelets help with blood clotting. As these blood cells. They used the K-medoids algorithm to extract WBCs from
cells are huge in number, traditional manual blood cell counting the image and granulometric analysis to separate the RBCs from
system using haemocytometer is highly time consuming and WBCs and then counted the number of cells using the labelling
erroneous and most of the cases accuracy vastly depends on the algorithm and a circular Hough transform (CHT). Sarrafzadeh
skills of a clinical laboratory analyst [3, 4]. Therefore, an et al. [11] proposed circlet transform to count RBCs on the grey-
automated process to count different blood cells from a smear scale image. They used iterative soft-thresholding method for
image will greatly facilitate the entire counting process. identification and counting purposes. Kaur et al. [12] proposed a
With the development of machine learning techniques, image method to count platelets automatically by applying a CHT in a
classification and object detection applications are becoming microscopic blood cells image. They used the size and shape
more robust and more accurate. As a result, machine learning features of platelets from the CHT in the counting process.
based methods are being applied in different fields. Particularly, Cruz et al. [1] presented an image processing system to count
deep learning methods are being applied in different medical appli- blood cells. They used hue, saturation, value thresholding
cations such as abnormality detection and localisation in chest method, and connected component labelling for the identification
X-rays [5], automatic segmentation of the left ventricle in cardiac and counting of blood cells. Acharjee et al. [3] proposed a
MRI [6], and detection of diabetic retinopathy in retinal fundus semi-automated process by applying a Hough transform to count
photographs [7]. Thus, it is worth to look into deep learning RBC by detecting their oval and biconcave shape. Lou et al. [4]
based methods that can be applied to identify and count the blood provided a method to automatically count RBCs using spectral
cells in the smear images. angle imaging and support vector machine (SVM). Zhao et al.
In this Letter, a deep learning based blood cell counting method [13] proposed an automatic identification and classification
has been proposed. We employ a deep learning based object system for WBCs using the convolutional neural network (CNN).
detection method to detect different blood cells. Among the Firstly, they detected WBCs from the microscopic images, and
state-of-the-arts object detection algorithms such as regions with then CNN was used to detect kinds of WBCs. Habibzadeh et al.
convolutional neural network (R-CNN) [8], you only look once [2] presented a system for classifying five different types of
(YOLO) [9], we chose YOLO framework which is about three WBCs. They used three classifiers, which include two different
times faster than Faster R-CNN with VGG-16 architecture [9]. SVMs and one CNN classifier. Habibzadeh et al. [14] employed
YOLO uses a single neural network to predict bounding boxes pre-trained CNNs, ResNet [16] and Inception Net [17], to count
and class probabilities directly from the full image in one evalu- WBCs from segmented images. The images were segmented
ation. We retrain YOLO framework to automatically identify employing colour space analysis. Xu et al. [15] employed patch
and count RBCs, WBCs, and platelets from blood smear images. size normalisation on pre-processed images and then applied
To improve the counting accuracy, a verification method has been CNN to classify RBC shapes from microscopy images of patients
developed to avoid repeated counting by the framework. of sickle cell disease.

Healthcare Technology Letters, 2019, Vol. 6, Iss. 4, pp. 103–108 103

doi: 10.1049/htl.2018.5098 This is an open access article published by the IET under the
Creative Commons Attribution License ([Link]
org/licenses/by/3.0/)
 different versions of it, we choose to use Tiny YOLO as it is the
fastest of all. Tiny YOLO uses 9 instead of 24 convolutional
layers other than that all the parameters are the same [9].

3.3. Training: The original implementation of the Tiny YOLO

configuration was trained for 20 different classes. To adopt it for
blood cells identification, we modify it for three classes
consisting of WBC, RBC, and platelets. Due to modifying the
class number, the number of filters in the final convolutional
layer in the CNN architecture is needed to be changed as well.
YOLO predicts five values along with class probabilities for each
anchor box. The values are the probability of having an object in
a grid cell, x and y coordinates of the object, height, and width of
the object. In our case, the number of anchor boxes is 5 as it will
provide better flexibility to put bounding boxes according to the
aspect ratio of the object [18]. The number of filters in the final
convolution layer, NF , can be computed from the number of
anchor boxes NA and number of classes NC by

NF = NA × (NC + 5). (1)

Fig. 1 Block diagram of automatic blood cells identification and counting
system Since NA is 5 and NC in 3 in the experiments, NF is found to be 40.
We use 300 annotated blood smear images for training and 60 for
testing. During training in each step, we record loss and moving
We propose a completely different approach that employs average loss. We record data for a total of 4500 steps and use
YOLO to detect all three types of blood cells simultaneously. two different learning rates. For steps 1–2500, we specify the
Our method does not require any greyscale conversion or binary learning rate 10−5 and for steps 2501–4500, the learning rate
segmentation. The whole process is fully automated, fast, and 10−7 . We found that, a lower learning rate at the later steps
accurate. enables better convergence. We recorded the weights and evaluated
the model after every 125 steps. Fig. 2 shows the learning curve of
3. Materials and method: Our goal is to use the object detection the YOLO algorithm for blood cell detection in terms of the loss
and classification algorithm YOLO to detect and count blood function. The value of the loss function is shown as is as well as
cells directly from smear image. We need to train the YOLO by the moving average of the loss function. It is seen from
framework with a modified configuration and annotated blood the figure that, the minimum moving average loss is found to be
cells training images. 8.8766 on step 3750 using the learning rate 10−7 . We use the
weights of this step for testing purpose.
3.1. Dataset: We use a publicly available dataset of annotated
blood cell images called Blood Cell Count Dataset (BCCD) 3.4. Proposed blood cell identification and counting method:
[BCCD: [Link] Originally Our proposed method is a machine learning approach where we
it has a total of 364 annotated smear images, but the dataset use YOLO algorithm for automatic identification and counting
has some crucial flaw. After splitting the dataset into training of blood cells. It includes a training model with a modified
(300) and testing (64) parts, we find that one annotation file in configuration where we change the final convolution layer for
the test set does not include any RBC, although the image three outputs, identification of blood cells with an appropriate
contains RBCs. Moreover, three annotations file exhibit very low
RBC than actual. So, we remove four fallacious files and the
total size of the test set becomes 60. For the validation set,
we randomly pick 60 training images with annotations.
This modified dataset can be downloaded from this GitHub
repository [Link]
Cell-Count-Dataset. To test our model on a different dataset, we
used the data from [11]. The dataset includes 100 images of
resolution 3246 × 2448 acquired by Nikon V1 camera mounted
on a Nikon ECLIPSE 50i microscope with a magnification of 100×.

3.2. YOLO: ‘You Only Look Once’ in short YOLO is a

state-of-the-art object detection classification algorithm [9].
It treats object detection as a regression problem. It requires only
one forward propagation pass through the network to make a fast
prediction for both image class and location. It resizes the image
by 448 × 448 and divides the entire image into a 7 × 7 grid cell
and each grid cell predicts for two bounding boxes and
confidence score for the boxes. If the centre of the object falls
into a grid cell that the grid cell is responsible for detecting
that object. The original implementation of the YOLO model as
a CNN evaluated on the PASCAL VOC dataset. Its network
architecture contains 24 convolutional layers and 2 fully
connected layers and inspired by the GoogLeNet. Among Fig. 2 Learning curve of the YOLO framework for blood cell identification

104 Healthcare Technology Letters, 2019, Vol. 6, Iss. 4, pp. 103–108

This is an open access article published by the IET under the doi: 10.1049/htl.2018.5098
Creative Commons Attribution License ([Link]
org/licenses/by/3.0/)
threshold, and count them from their labels. We choose
x − x 
the threshold value by calculating the average absolute error r= 2 1
(3)
2
between ground truths and our estimation at difference threshold
value and realise the appropriate threshold for each type of cell
that gives a minimum average absolute error in the validation We count cells using their label. The modified YOLO returns three
dataset. Our proposed method does not misinterpret among kinds of labels ‘RBC’, ‘WBC’, and ‘Platelets’ depending on the
cells such as identifying RBC as WBC or platelets as RBC and detected cell. The total number of RBC in a smear image will be
so on. In some cases, it double count platelets. We resolve this the total number of labels containing ‘RBC’, the total number
by using K-nearest neighbour (KNN) and intersection over union of WBC will be the total number of labels containing ‘WBC’
(IOU) in each platelet. Overall, our proposed method is fast and and so on.
accurate in the identification and counting of blood cells. In some cases, our models provide two different detections for
The steps of the proposed method are described in Algorithm 1 a single platelet. We observed that the reason is the detection
(see Fig. 3). of the same platelet from two consecutive grid cells, and thus
We can get four parameters from the YOLO model for each the same platelet is counted twice. To avoid this double counting
detected cell. They are the label of the cell, the confidence of problem, we apply the KNN algorithm in each platelet and
being that cell, top left corner position, and bottom right corner determine its closest platelet and then using the intersection of
position. To show which of the cells are detected in the blood union (IOU) between two platelets we calculate their extent of
smear image, we have two choices. Using the top left and bottom overlap. Using empirical observations, we allow 10% of the
right corner coordinates, we can put a rectangular bounding box overlap between platelet and its closest platelet. If the overlap is
that encloses each detected cell. However, blood cells are not larger than that, we ignore that cell as double count to get rid of
rectangular rather close to circular in shape, and rectangular spurious counting. Fig. 4 shows such a case where a platelet is
boxes occupy much redundant space than it requires. So, we detected twice by the YOLO algorithm. Using the proposed KNN
place circular bounding boxes to enclose each cell, and that requires and IOU based technique, this double detection problem has been
the conversion of the top left and bottom right coordinates to radius removed.
and centre of the circle.
Given the top left and bottom right coordinates are (x1 , y1 ) and 4. Experiments and results: With the proposed method, we
(x2 , y2 ), the centre point C and the radius r of the circle that encloses automatically identify and count RBCs, WBCs, and platelets.
the cell can be calculated by We test our model using a test dataset of 60 images where the
ground truths are known. First, we use our model to count

x + x y + y 
the different cells in the validation dataset with different
C= 1 2
, 1 2
(2)
2 2 confidence threshold. It is noted that the threshold plays an
instrumental role in YOLO as it uses this threshold to predict
each grid cell, not for the whole image. Grid cell containing no
blood cell has low confidence. So, we can get rid of redundant
and spurious predictions by choosing appropriate confidence
threshold.
We calculate the average absolute error between ground truths
and the estimated number of cells in the validation dataset. With
different confidence threshold, we realise the minimum average

Fig. 4 Example blood smear image showing

a Counting of the same platelet twice
b Discarding spurious prediction using the proposed method

Table 1 Average absolute error ground truths and estimated number of

RBCs, WBCs, and platelets at a different confidence threshold

Threshold, % RBC WBC Platelets

20 5.650 0.083 0.217

25 4.417 0.050 0.083
30 3.450 0.033 0.083
35 2.750 0.017 0.083
40 2.500 0.050 0.083
45 2.183 0.100 0.100
50 2.133 0.150 0.100
55 2.083 0.200 0.117
60 2.100 0.333 0.150
Fig. 3 Algorithm 1: automatic blood cell identification and counting

Healthcare Technology Letters, 2019, Vol. 6, Iss. 4, pp. 103–108 105

doi: 10.1049/htl.2018.5098 This is an open access article published by the IET under the
Creative Commons Attribution License ([Link]
org/licenses/by/3.0/)
Table 2 Accuracy of counting RBC, WBC, and platelets employing the absolute error value for each type of cell and choose those confi-
proposed method dence values in the identification process of blood cells. The error
is computed using
RBCs WBCs Platelets
1 N
1cell = |x(i) − x(i)
estimated | (4)
ground truths 792 61 55 N i=1 groundtruths
estimated 823 53 53
accuracy, % 96.09 86.89 96.36 where cell indicates the type of cells (RBC, WBC, or platelets), N is
the size of validation dataset (in our experiment it is 60), x is
the number of cells, 1 is the average absolute error value for the
particular cell. The computed error values are shown in Table 1.
It is seen from the table that for counting RBCs, we can employ
a nominal threshold of 0.55. However, for WBC and platelets,
the threshold is found to be much lower (0.35 and 0.25 in our
experiments). Thus, the appropriate thresholds for each type of
cell are selected as follows:

† RBC: confidence threshold of 55%.

† WBC: confidence threshold of 35%.
† Platelets: confidence threshold of 25%.

Fig. 5 Comparison of the ground truth and predicted blood cell identifica- Then, we have calculated the accuracy from the total number
tion output of ground truths cells and the total number of estimated cells
a Ground truth labels of cells in a smear image in the test dataset. With a confidence threshold of 55% for RBC,
b Automatically estimated labels of cells by our model we achieved 96.09% accuracy for RBC. Total estimated numbers
of cells of different types with accuracy calculated at their appropri-
ate confidence threshold value are presented in Table 2. It may seem
that the proposed algorithm is counting more or extra RBCs that are
not in the images. However, we would like to note that the ground
truth labels were not present for some of the RBCs that are at the
edge of the image. The YOLO algorithm can detect these RBCs,
and thus, the RBC count is high.
To visualise, the output of the proposed method concerning
ground truth, a sample smear image from the test set is shown in
Fig. 5. It is seen from the figure that all the WBC and platelets
are detected without error. The method has missed one RBC in
the middle, whereas detected another RBC from the edge of the
image which is not present in the ground truth.

4.1. Experiments with other CNN architectures: YOLO has a

built-in CNN architecture for classification which is inspired by
GoogLeNet architecture. Besides training blood cell detection
model with YOLO’s own CNN, we have experimented with other
popular CNN architectures. We use VGG-16 [19], ResNet50
[16], InceptionV3 [17], and MobileNet [20] CNN architectures
with the YOLO algorithm by replacing it’s built-in CNN. To train
YOLO using these networks in the backend, we have split our
training dataset into two parts. First 250 images with annotations
are used for training purposes and rest 50 s are used for
Fig. 6 CNN models loss curves with YOLO algorithm along with their val- validation purposes. For all the networks training loss curves
idation mAP along with validation mean average precision (mAP) values are

Table 3 Accuracy of detecting different cells using different CNN architecture with YOLO algorithm

RBC WBC Platelet mAP Execution time, ms

Ground truths 792 61 55

Tiny YOLO estimated 823 53 53 0.6236 60

accuracy, % 96.09 86.89 96.36
VGG-16 estimated 1006 61 60 0.7132 106
accuracy, % 72.98 100 90.91
ResNet50 estimated 952 58 48 0.7437 118
accuracy, % 79.80 95.08 87.27
InceptionV3 estimated 889 61 57 0.6826 130
accuracy 87.75 100 96.36
MobileNet estimated 588 57 46 0.5207 84
accuracy, % 74.24 93.44 83.64

106 Healthcare Technology Letters, 2019, Vol. 6, Iss. 4, pp. 103–108

This is an open access article published by the IET under the doi: 10.1049/htl.2018.5098
Creative Commons Attribution License ([Link]
org/licenses/by/3.0/)
 identifies the RBCs, WBCs, and platelets with satisfactory
performance.

5. Conclusion: In this Letter, a machine learning approach to

automatically identify and count blood cells from a smear image
based on YOLO algorithm is presented. To improve accuracy,
the method employed KNN and IOU based method to remove
multiple counting of the same object. Our proposed method is
evaluated on publicly available datasets. It is observed for test
dataset that, our method accurately identifies RBCs, WBCs, and
Platelets. It is seen that our method can accurately count even
some of the cells that are not labelled in the dataset. Different
neural network models have also been tried in the YOLO
back-end, and it has been observed that different models can
provide the best accuracy on different cells. Even though different
models with different depths have been tried, it is observed that
the method is considerably fast for counting and marking the
smear images. The proposed method has also been tested on a
different dataset of smear images, where it has performed
satisfactorily. With the accuracy and the detection performance of
the proposed method, it can be said that, the method has the
potential to ease up the manual blood cell identification and
counting process.

6. Funding and declaration of interest: Conflict of interest: none

declared.

7 References
[1] Cruz D., Jennifer C., Valiente L.C., ET AL.: ‘Determination of blood
components (WBCs, RBCs, and platelets) count in microscopic
images using image processing and analysis’. Proc. Int. Conf.
Fig. 7 Blood cells detection in an image from the dataset [11] on Humanoid, Nanotechnology, Information Technology,
a Image divided into 3 × 3 grids Communication and Control, Environment and Management
b Combined output (HNICEM), December 2017
[2] Habibzadeh M., KrzyÅijak A., Fevens T.: ‘White blood cell differen-
tial counts using convolutional neural networks for low resolution
shown in Fig. 6. It is seen from the figure that, InceptionV3 and images’. Proc. Int. Conf. on Artificial Intelligence and Soft
ResNet50 achieves the lowest error. Computing, 2013, pp. 263–274
The accuracy of counting RBC, WBC, and platelets on the test [3] Acharjee S., Chakrabartty S., Alam M.I., ET AL.: ‘A semiautomated
approach using GUI for the detection of red blood cells’. Proc. Int.
set is shown in Table 3. It is seen from the table that, the highest Conf. on Electrical, Electronics, and Optimization Techniques,
accuracy for counting RBCs and platelets is found for the Tiny 2016, pp. 525–529
YOLO architecture, which achieved 96.09% accuracy in counting [4] Lou J., Zhou M., Li Q., ET AL.: ‘An automatic red blood cell counting
RBCs and 96.36% accuracy in counting platelets. On the other method based on spectral images’. Proc. Int. Congress on Image and
hand, VGG-16 and InceptionV3 both have achieved the highest Signal Processing, BioMedical Engineering and Informatics, October
accuracy in counting WBCs (100%). Thus, different models can 2016, pp. 1391–1396
[5] Islam M.T., Aowal M.A., Minhaz A.T., ET AL.: ‘Abnormality
achieve the highest accuracy for different cells. We have also com- detection and localization in chest X-rays using deep con-
puted the mean average precision (mAP) of these CNN architec- volutional neural networks’, arXiv preprint arXiv:1705.09850v3,
tures and their execution time per test image listed in Table 3. 2017
It is seen from the table that ResNet50 achieves highest mAP [6] Avendi M.R., Kheradvar A., Jafarkhani H.: ‘A combined
value. Our proposed method is extremely fast requiring less than deep-learning and deformable-model approach to fully automatic seg-
a second even for deeper networks. We calculate the forward mentation of the left ventricle in cardiac MRI’, Med. Image Anal.,
2016, 30, pp. 108–119
propagation time for each of the test images on a computer with [7] Gulshan V., Peng L., Coram M., ET AL.: ‘Development and validation
Intel Core i5 6500 with 8 GB memory and Nvidia GTX1050 Ti of a deep learning algorithm for detection of diabetic retinopathy in
GPU with 4 GB memory. Estimation time is reported by computing retinal fundus photographs’, J. Am. Med. Assoc., 2016, 316, (22),
the average to get the final execution time. pp. 2402–2410
[8] Ren S., He K., Girshick R., ET AL.: ‘Faster R-CNN: towards teal-time
object detection with region proposal networks’, IEEE Trans. Pattern
4.2. Testing using a different dataset: To observe, whether the Anal. Mach. Intell., 2017, 39, (6), pp. 1137–1149
[9] Redmon J., Divvala S., Girshick R., ET AL.: ‘You only look once:
trained model is database dependent, we have also tested our unified, real-time object detection’. IEEE Conf. on Computer
model using blood smear images from another dataset [11]. The Vision and Pattern Recognition, December 2016
images of this dataset were of higher resolution, and thus the [10] Acharya V., Kumar P.: ‘Identification and red blood cell automated
images were divided into grids to create sub-images, and each of counting from blood smear images using computer-aided system’,
the sub-images has been processed individually in the proposed Med. Biol. Eng. Comput., 2018, 56, (3), pp. 483–489
pipeline. The detection and counting from each of the sub-images [11] Sarrafzadeh O., Dehnavi A.M., Rabbani H., ET AL.: ‘Circlet based
framework for red blood cells segmentation and counting’. IEEE
are then projected back on the original image. A blood smear
Workshop on Signal Processing Systems, December 2015
image which has been divided into a 3 × 3 grid, which causes 9 [12] Kaur P., Sharma V., Garg N.: ‘Platelet count using image processing’.
sub-images to be processed in the proposed method, is shown in Int. Conf. on Computing for Sustainable Global Development
Fig. 7. It is seen from the figure that, the output image correctly (INDIACom), March 2016

Healthcare Technology Letters, 2019, Vol. 6, Iss. 4, pp. 103–108 107

doi: 10.1049/htl.2018.5098 This is an open access article published by the IET under the
Creative Commons Attribution License ([Link]
org/licenses/by/3.0/)
[13] Zhao J., Zhang M., Zhou Z., ET AL.: ‘Automatic detection and [17] Szegedy C., Vanhoucke V., Ioffe S., ET AL.: ‘Rethinking the inception
classification of leukocytes using convolutional neural networks’, architecture for computer vision’. IEEE Conf. on Computer Vision
Med. Biol. Eng. Comput., 2017, 55, (8), pp. 1287–1301 and Pattern Recognition, December 2016
[14] Habibzadeh M., Jannesari M., Rezaei Z., ET AL.: ‘Automatic [18] Redmond J., Ali F.: ‘YOLO9000: better, faster, stronger’, arXiv
white blood cell classification using pre-trained deep learning models: preprint arXiv:1612.08242, 2016
ResNet and inception’. AIP Conf. Proc., 2017, vol. 1883, pp. 1–8 [19] Simonyan K., Zisserman A.: ‘Very deep convolutional networks
[15] Xu M., Papageorgiou D.P., Abidi S.Z., ET AL.: ‘A deep convol- for large-scale image recognition’. Proc. Int. Conf. Learning
utional neural network for classification of red blood cells in sickle Representations, San Diego, 2015, pp. 1–14
cell anemia’, PLoS Comput. Biol., 2017, 13, (10), pp. 1–27 [20] Howard A.G., Zhu M., Kalenichenko B.C.D., ET AL.:
[16] He K., Zhang X., Ren S., ET AL.: ‘Deep residual learning for image ‘Mobilenets: efficient convolutional neural networks
recognition’. IEEE Conf. on Computer Vision and Pattern for mobile vision applications’, arXiv preprint arXiv:1704.04861,
Recognition, December 2016 2017

108 Healthcare Technology Letters, 2019, Vol. 6, Iss. 4, pp. 103–108

This is an open access article published by the IET under the doi: 10.1049/htl.2018.5098
Creative Commons Attribution License ([Link]
org/licenses/by/3.0/)