Micron 83 (2016) 32–41
Contents lists available at ScienceDirect
Micron journal homepage: www.elsevier.com/locate/micron
Healthy and unhealthy red blood cell detection in human blood smears using neural networks Hany A. Elsalamony a,b,∗ a b
Mathematics Department, Faculty of Science, Helwan University, Cairo, Egypt Computer Science & Information Department, Arts & Science College, Sattam University, Saudi Arabia
a r t i c l e
i n f o
Article history: Received 1 December 2015 Received in revised form 26 January 2016 Accepted 27 January 2016 Available online 1 February 2016 Keywords: Healthy/unhealthy RBC detection and counting Circular hough transforms Segmentation Neural network
a b s t r a c t One of the most common diseases that affect human red blood cells (RBCs) is anaemia. To diagnose anaemia, the following methods are typically employed: an identification process that is based on measuring the level of haemoglobin and the classification of RBCs based on a microscopic examination in blood smears. This paper presents a proposed algorithm for detecting and counting three types of anaemia-infected red blood cells in a microscopic coloured image using circular Hough transform and morphological tools. Anaemia cells include sickle, elliptocytosis, microsite cells and cells with unknown shapes. Additionally, the resulting data from the detection process have been analysed by a prevalent data analysis technique: the neural network. The experimental results for this model have demonstrated high accuracy for analysing healthy/unhealthy cells. This algorithm has achieved a maximum detection of approximately 97.8% of all cells in 21 microscopic images. Effectiveness rates of 100%, 98%, 100%, and 99.3% have been achieved using neural networks for sickle cells, elliptocytosis cells, microsite cells and cells with unknown shapes, respectively. © 2016 Published by Elsevier Ltd.
1. Introduction The general components of human blood are plasma, white blood cells (WBCs), red blood cells (RBCs) and platelets. RBCs comprise approximately 40% of blood volume. WBCs are smaller in volume but larger in size than RBCs. Plasma is the fluid component that contains melted salts and proteins. Platelet cells are similar particles but are smaller than WBCs and RBCs (Xia and Wu, 2015; Biradar et al., 2015; Deligiannidis and Arabnia, 2014). Anaemia is a type of red blood cell disorder that is usually caused by a lack of mineral iron in the blood. The human body needs iron to produce the iron-rich protein haemoglobin, which helps red blood cells carry oxygen from the lungs to the remainder of the body (Elsalamony, 2014; Das et al., 2013). This disease occurs when the blood has a lower than normal number of red blood cells (RBCs) or an insufficient amount of haemoglobin. RBCs are located inside the large bones of the body in the spongy marrow. The main function of marrow is to renew red blood cells, which continuously replaces old red blood cells. Normal RBCs die after they have lived in the bloodstream for 120 days. Their jobs include carrying oxygen and
∗ Correspondence to: Mathematics Department, Faculty of Science, Helwan University, Cairo, Egypt. E-mail addresses: h [email protected], [email protected] http://dx.doi.org/10.1016/j.micron.2016.01.008 0968-4328/© 2016 Published by Elsevier Ltd.
removing carbon dioxide (a waste product) from the body (Lam, 2015). RBCs are disc-shaped and can easily move through blood vessels. Elliptocytosis is a well-known type of anaemia. Historically, this disease was described in 1904 and recognized as a hereditary condition in 1932. The medical determination of hereditary elliptocytosis is difficult. The incidence of this disease ranges between three and five cases per 10,000 in the USA, whereas an estimated 60–150 cases per 10,000 of African and Mediterranean natives and 1500–2000 per 10,000 cases of Malayan natives have been documented (Lee and Chen, 2014). In sickle-cell anaemia, which is a serious disorder, the body creates a crescent shape of red blood cells. These sickle cells contain abnormal haemoglobin, which is referred to as sickle haemoglobin or haemoglobin S; it helps cells to develop a crescent shape. The absence of a polar amino acid encourages the noncovalent combination of haemoglobin in a low-oxygen environment, which distorts the red blood cells into a sickle shape and decreases their elasticity. Biochemically, the low-oxygen environment causes a chain of neighbouring haemoglobin molecules to hook together and block blood flow in the blood vessels of the limbs and organs, which become rigid and polymerized. Low blood flow can cause pain, organ damage, and increase the probability of disease. These cells fail to return to their normal shape when oxygen is restored and fail to deform as they pass through nar-
H.A. Elsalamony / Micron 83 (2016) 32–41
row vessels, which causes blockages in the capillaries ([2015b). Abnormal sickle cells usually die after approximately ten to 20 days. Bone marrow cannot renew red blood cells fast enough to replace dying red blood cells (Thirusittampalam et al., 2013). No cure exists for sickle cell anaemia. However, treatments can help to release pain and improve the complications of this disease. In addition, sickle-cell anaemia is common in people whose families derive from Mediterranean countries, Africa, South, or Central America, especially Panama, the Caribbean islands, Saudi Arabia, and India. In the United States, the disease derived from 70,000 to 100,000 people, primarily African Americans. The diagnosis of sickle cell anaemia is dependent on blood test analyses that can detect sickle cells (2015a). In the same context, microsites consist of small red blood cells. The most common causes of microcytic anaemia are iron deficiency and the thalassemia trait. The ability to distinguish microsites is very important for providing genetic counselling and preventing unnecessary and damaging iron therapy in thalassemia carriers. One of the simplest and most powerful discriminant functions is the ratio of microcytic cells to normal red blood cells (Urrechaga et al., 2015). The average size of a normal red blood cell is a mean corpuscular volume (MCV) in the range of 80–100 FL; smaller cells (90% total discrimination accuracy) perceived by FLDA, and PCA detected three critical principal components, which cumulatively elucidated 98.4% of the aggregate variance. In 2014, Lee and Chen (2014) introduced a neural network model with a classifier, which utilize the visual information obtained from the images of red blood cells images to determine if a red blood cell is standard or odd. They clustered the visual components into two essential parts—shape cluster groups and texture cluster groups—depending on the component properties. The input feature clusters were considered using parallel and course construction with input various data layers. Their trial results demonstrated imperative change and exactness in the proposed structure, which was distinguished from the single information layer classifier with element selection algorithms. In 2015, Yi et al. (2015) proposed a three-dimensional characterization strategy for consequently identifying the morpho-
34
H.A. Elsalamony / Micron 83 (2016) 32–41
Fig. 1. Difference in shape among normal RBCs.
logically ordinary RBCs in images of different human RBCs that were acquired by off-pivot computerized holographic microscopy (DHM). The 3D images of RBCs were recorded by DHM, and the stage images of various RBCs were subsequently reproduced by computational numerical calculation. They performed three runs of the typical RBC shapes—stomatocyte, discocyte, and echinocyte—for training and testing. The unusual RBC shapes were characterized as a fourth class. They obtained ten components, which were separated from every RBC, to fragment the reproduced stage images as a watershed transform, which include the projected surface region, average phase value, mean corpuscular haemoglobin, perimeter, mean corpuscular haemoglobin surface density, circularity, mean phase of centre part, sphericity coefficient, elongation, and pallor. They applied a principal component analysis algorithm to decrease the measurement number of variables and increase the Gaussian mixture densities by utilizing the projected data and the initial eight principal components. Therefore, the Gaussian mixtures were employed to plan the discriminant capacities based on Bayesian decision theory. Their experimental results demonstrated that the proposed strategy can achieve satisfactory results for ascertaining the rate of each average ordinary RBC shape in reconstructed stage images of numerous RBCs.
and (2)—to reduce errors in detection operation and ensure a final diagnosis of anaemia. The final stage of the proposed algorithm involves displaying the input image with all contoured healthy (green) and unhealthy (other colours) RBCs, in which all tested images are set to the same distance (focal length) from the microscope optics to control all measures. Via this operation, the data variables for each cell are measured, counted, and diagnosed for the majority of types in a smear. These measures serve as the input variables to train NN, whereas the output targets are measured based on the solidity (SC ), which is calculated by dividing the area by its convex area for all separate cells, as shown in Eq. (1), and the ratio (RC ), which is the ratio that is calculated by dividing the area of cells to the perimeter of cells, as shown in Eq. (2) SC =
areac (convexareac )
where areac is the area of each cell for all types of detected cells and convexareac represents the convex area of each cell for all types of detected healthy/unhealthy cells. RC =
3. The proposed algorithm The goal of this paper is to detect healthy and unhealthy blood cells in a coloured microscopic image of a human blood smear (Xia and Wu, 2015). The proposed algorithm is illustrated in Fig. 2. Fig. 2 presents an application of CHT with morphological functions on bright and dark intensity cells to detect and count healthy/unhealthy blood cells. CHT was employed to detect RBCs even if they were hidden, on image borders or overlapped. Then, watershed and morphological functions were employed to enhance and separate overlapped cells during the segmentation process. In CHT, many operations, such as cell polarity, which indicates whether circulating blood cells are brighter or darker than the background, and a computation method (two-stage), which calculates the accumulator array of CHT based on computing radial histograms, re apply estimated cell centres and the image information (Elsalamony). Another operation is the sensitivity factor, which is the sensor of the accumulator array in CHT. The detection process includes weak and partially hidden or overlapped cells. However, higher values of sensitivity increase the risk of false detection. The edge gradient threshold is the last operation that is performed; cells generally exhibit a darker interior (nuclei) that is surrounded by an outlying bright halo. The edge gradient threshold is very useful for determining edge pixels in images of both unhealthy and healthy blood cells based on their contrast, which are easily detected by establishing a lower threshold value. Fewer cells of weak edges are detected by increasing the value of the threshold (Deligiannidis and Arabnia, 2014). The NN has been applied to test and analyze the performance of the proposed algorithm for diagnosing and determining whether a patient has one or more types of anaemia. The classification is dependent on the data variables of the detected cells—area, convex area, perimeter, eccentricity, solidity, and ratio (refer to Eqs. (1)
(1)
4(areac )
(2)
2
(perimeterc )
The perimeterc represents the perimeter of each cell for all types of detected healthy/unhealthy cells. The target TSC is an array that consists of the binary numbers (1) and (0) for the conditions of solidity and eccentricity to classify sickle cell anaemia, as shown in Eq. (3).
TSC =
1 if SC < 0.72 or Eccentricity > 0.9 0
(3)
Otherwise
If any of the solidity values SC is less than 0.72 or the eccentricity is greater than 0.9, then the diagnosis is sickle cell anaemia. Otherwise, TSC assumes a value of zero with a diagnosis of healthy/unhealthy cells; however, the type anaemia is not sickle cell anaemia. The same conditions apply to target TRC , which is an array that is constructed in the conditions for ratio to classify elliptocytosis, as shown in Eq. (4).
TRC =
1 if 0.7 ≤ RC ≤ 0.9 0 Otherwise
(4)
when TRC is 1, the ratio RC ranges between 0.7 and 0.9, which indicates that the unhealthy cell belongs to elliptocytosis. When RC is less than 0.7 or greater than 0.9, then TRC is zero, which renders the decision that the cell is not derived from elliptocytosis and may be a healthy or other type of unhealthy cells. To detect microsites, Eq. (5) gives the condition that MC is 1 when the eccentricity exceeds 0.45 logical AND RC exceeds 0.9; otherwise, MC is zero.
MC =
1 if Eccentricity > 0.45 andRC > 0.9 0
Otherwise
(5)
H.A. Elsalamony / Micron 83 (2016) 32–41
35
Fig. 2. The main points of the proposed algorithm.
In Eq. (6), UC is an array that has been constructed in reverse conditions or conditions that complement all previous conditions in TSC , TRC and MC.
UC =
1 if Eccentricity > 0.9and SC ≥ 0.72 0
(6)
Otherwise
The target UC of unknown shapes is defined by 1 if the values of RC are less than 0.7, if logically and eccentricity are less than or equal to 0.9, and if SC is greater than or equal to 0.72. However, it assumes the value 0 when the value satisfies the conditions of TSC , TRC and MC ; then, the diagnosis is either healthy, sickle, elliptocytosis, or microsite. Table 1 lists values of variables according to healthy/unhealthy cells. The maximum and minimum values of the variables are listed according to the previously mentioned healthy/unhealthy cells. The minimum value of healthy cells is zero, which indicates that the cell is perfectly rounded or appears in a microscope in a complete position. Microsites are rounded with lower eccentricity values but smaller values than healthy cells. The back propagation neural network has trained and tested six variables as an input layer, ten neurons in the hidden layer and one neuron in the output layer. Three types of samples are applied: training samples, validation samples, and testing samples. The training samples represent 70% of all samples in the network during the training process in the input variables, and the network was modified according to the errors. Only 15% of the samples were employed in the validation to measure network generalization and to pause training when generalization stops improving. The remaining 15% of all samples of cells have been introduced as a testing sample, which has no effect on training and provides an independent measure of network performance during and after training. In addition, the mean square error (MSE), which is defined as the average squared difference between outputs and targets, was applied; lower values are better, and a value of zero indicates no error. Table 2 provides data on formed healthy/unhealthy cells. The range of all variables are similar, whereas only the output (TSC , TRC , MC , and UC ) variable has a binary value, such as 1, to exactly defined cells, and the remaining variables have a value of zero. The execution of every classification model has been assessed utilizing three factual measures: classification accuracy, sensitivity, and specificity. These benchmarks include the true positive (TP), true negative (TN), false positive (FP) and false negative (FN). A true positive choice occurs when the positive expectation of the classifier corresponds with a positive forecast of the previous segmentation. A true negative choice occurs when both the classifier and the division proposes the nonattendance of a positive expectation. A false positive choice occurs when the framework marks a healthy cell (positive expectation) as an unhealthy cell. A false negative choice occurs when the framework marks a negative (unhealthy) cell as a positive cell. In addition, the classification
accuracy has been characterized as a proportion of effectively ordered cells and is equivalent to all TP and TN that were isolated by the aggregate number of RBCs (N) (Bishop). Accuracy =
TP + TN N
(7)
The sensitivity alludes to the rate of correctly classified positive cells and is equivalent to the TP separated by the aggregate of TP and FN. Sensitivity =
TP TP + FN
(8)
Specificity refers to the rate of effectively arranged negative cells and is equivalent to the proportion of TN to the total TN and FP (Bishop). Specificity =
TN TN + FP
(9)
4. The experimental results The experimental results are displayed in two forms: the first form is the detection of healthy/unhealthy RBCs and the counting of these cells. The second form is the analysis and classification of the exported cells data to verify the agreement with the detection process using the most famous classification model in data analysis—the NN (Chaudhuri and Bhattacharya, 2000; Bishop, 2006). In the detection process, blood cells were obtained from subjects with anaemia, and the algorithm was employed using 21 selected images with a size of 504 × 700 for ten samples with 40× magnification to accurately detect the anaemia type—sickle, elliptocytosis, microsites, unknown shapes. Cells were thinly smeared onto glass magnifying lens slides to permit singular cell sorts to be recognized by utilizing the proposed calculation. Giemsa was utilized as a part of blood arrangements and stains: red blood cells (pink), and platelets and white blood cells (maroon).All cells (healthy/unhealthy) were detected using CHT, watershed, and morphological techniques for enhancement. In the same context, CHT was applied in the conditions of cell polarity to determine all dark and bright cells according to their intensity. Two-stage techniques were subsequently employed to compute the accumulator array of CHT. The sensitivity of this accumulator array in the proposed algorithm is 0.97 for brightness and 0.90 for dark cells, and an edge gradient threshold of 0.2 for the detection of fewer cells with weak edges. These conditions help CHT to detect the majority of the healthy RBCs that positioned singular, overlapped, and even attached to other unhealthy cells. Fig. 3(a) shows an example image of a human blood smear in, and some of the proposed algorithm steps are shown in (b) and (c). According to this image of the detection process, the number of healthy cells is 109 of 160, which is the total number of all detected cells (healthy/unhealthy). The remaining number of cells (51) may be considered as detection errors, unhealthy cells (sickle,
36
H.A. Elsalamony / Micron 83 (2016) 32–41
Table 1 Variables according to healthy/unhealthy cells. Variables
Healthy cells Unhealthy cells
Area
Convex area
Perimeter
Max
Min
Max
Min
Max
Min
Eccentricity Max
Min
Solidity Max
Min
Ratio Max
Min
797 869
539 158
821 1145
554 178
104.5685 150.3675
91.1127 51.1127
0.756632 0.97553
0 0.422419
0.97914 0.977337
0.940781 0.506494
0.929457 0.945956
0.815907 0.278963
Table 2 Blood cell variables. #
Variables
Type
Domain
1 2 3 4 5 6 7 8 9 10
Convex area Perimeter Eccentricity Areas Solidity Ratio TSC TRC MC UC
Ranged Ranged Ranged Ranged Ranged Ranged Binary Binary Binary Binary
178:1145 51.1127:150.3675 0:0.97553 158:869 0.506494:0.97914 0.278963:0.945956 (1 for sickle cell, 0 for other healthy/unhealthy cells) (1 for elliptocytosis, 0 for other healthy/unhealthy cells) (1 for microsites, 0 for other healthy/unhealthy cells) (1 for unknown shapes, 0 for other healthy/unhealthy cells)
Fig. 3. (a) The original image, (b) makes a small mark on every cell’s centroid, and (c) the final detected healthy cells.
elliptocytosis, microsites, and unknown), platelets, or even WBCs. Platelets and WBCs have been neglected (if they appeared in image) as the algorithm has only concentrated on healthy/unhealthy RBCs (Elsalamony). Additionally, the algorithm has determined 153 of the detected blood cells, without detection of unhealthy cells on the image borders, by approximately 97.8% according to the image in Fig. 3(a). Consequently, the next step is trying to classify these 51 cells into several types of unhealthy cells. All cells have strange shapes (crescent, elliptic, small RBCs, platelets, and unknown shapes); they are discovered and selected as unhealthy cells by the previous steps of healthy cell detection. Fig. 4 illustrates the detection of the majority of unhealthy cells without any cells in the image borders. In Fig. 4(a), a coloured segmented image for all unhealthy shapes with noise-like platelets. The sickle cells, elliptocytosis, microsites and unknown shapes have been detected in different contour colours, and the platelets have been neglected. The blue line contours represent the elliptocytosis cells, the cyan line represents microsites, the red line represents sickle cells (crescent or initiated to be crescent), whereas the yellow line represents unknown cell shapes. In the same context, of the 51 cells, the deformed cells include 15 sickle cells or initiated to be, 18 elliptocytosis cells, two of microsites and seven unknownshaped cells. The remaining seven unhealthy cells, which are not detected but may appear on the neglected image borders, include unknown detected cells that contain undefined or unknown composite shapes, or waste or unknown blood disease. In Fig. 4(b), the final detection of healthy cells are denoted by green lines and unhealthy cells are represented by a variety of coloured lines.
Experimentally, NN consists of six input variables and ten neurons in one hidden layer and one output layer. Three networks have trained with three targets—TSC , TRC and MC —for the detection of sickle, elliptocytosis and microsite cells; the fourth target has trained for unknown shapes with the target UC . The first NN is concerned with the detection of sickle cells, which is succeeded in 14 s after 29 iterations of 1000 as maximum iterations of the epoch, the performance 1.11e−07 , the gradient 7.67e−07 , and zero validation checks, as shown in Fig. 5(a). In Fig. 5(b), the best validation performance is shown as 3.6828e−09 at epoch 29, where the training is denoted by a blue line, the validation is denoted by a green line, and the test is denoted by a red line. The mean square error for training, validation, and testing is 1.10875e−07 , 3.6828e−09 , and 1.8058e−13 , respectively. The confusion matrices for the training, validation, and testing processes of the first NN are shown in (c). The predictions of the NN model are compared with the original classes of the target TSC to recognize the estimations of true positives, true negatives, false positives, and false negatives. These qualities have been processed to develop the confusion matrix, in which every cell contains the quantity of the cases that are arranged for the relating mix of craved and genuine classifier yields of 100%. An accuracy, sensitivity, and specificity of 100% is achieved. In the second NN model, the detection of elliptocytosis cells succeeded in 0 s after 56 iterations of 1000 as maximum iterations of the epoch, a performance of 0.00908, a gradient of 0.0643, and six validation checks, as shown in Fig. 6(a). In the same context, the best validation performance is shown as 0.029644 at epoch 50, where the training is denoted by a blue line, the validation is denoted by a green line, and the test is represented by the red line of (b). The mean square error of training, validation,
H.A. Elsalamony / Micron 83 (2016) 32–41
37
Fig. 4. (a) All contoured unhealthy cells, and (b) detection of all cells. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
Fig. 5. (a) The first NN Sickle cells, (b) the best validation performance, and the confusion matrices in (c). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
and testing are 1.02479e−02 , 2.9644e−02 , and 1.87565e−02 , respectively. The confusion matrices of the second NN with the target TRC are displayed in (c); 98% performance is achieved. Accordingly, accuracy, sensitivity, and specificity have achieved 99.1%, 96.4%, and 100%, respectively, due to the success of the training samples. They have achieved 95.7%, 75%, and 100% due to the success of the testing samples. The microsites NN has succeeded in 0 s after 23 iterations of 1000 as maximum iterations of the epoch, performance 6.72e−05 , gradi-
ent 0.000403, and six validation checks, as shown in Fig. 7(a). The best validation performance is 7.4256e−06 at epoch 17, as shown in Fig. 7(b). The mean square error for training, validation, and testing processes are 3.2056e−04 , 7.42563e−06 , and 1.62962e−03 , respectively. The confusion matrices of the third NN with the target MC are shown in (c); 100% performance is achieved. Accuracy, sensitivity, and specificity have achieved 100% due to the success of the training samples; they have achieved the same success of the testing samples.
38
H.A. Elsalamony / Micron 83 (2016) 32–41
Fig. 6. (a) The back propagation of the second NN elliptocytosis cells, (b) the best validation performance, and the confusion matrices in (c). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
The fourth NN is concerned with the detection of unknownshaped cells. It has succeeded, in 15 s, after 39 iterations of 1000 as maximum iterations of the epoch, performance of 0.000371, gradient of 0.00229, and six validation checks, as shown Fig. 8(a). Consequentially, the best validation performance is 0.0053972 at epoch 33, with the same coloured lines as shown in (b). The mean square error for training, validation, and testing are 1.60196e−03 , 5.39719e−03 , and 1.74046e−02 , respectively. The confusion matrices of the fourth NN with the target UC are shown in (c); 99.3%
performance is achieved. Accuracy, sensitivity, and specificity have achieved 100% due to the success of the training samples; however, they have achieved 95.7%, 100%, and 95.5% due to the success of the testing samples. The proposed algorithm is very effective in the diagnosing, counting, and detection of the healthy/unhealthy cells of a human blood smear.
H.A. Elsalamony / Micron 83 (2016) 32–41
39
Fig. 7. (a) The back propagation of third NN microsite cells, (b) the best validation performance, and the confusion matrices in (c).
5. Materials and devices Matlab-2013a was employed to construct the algorithm; a 32bit operating system with Windows 8, an Intel® CoreTM 2Duo CPU T5550 processor, 1.83 GHz and 2.50 GB RAM. The optical Nikon microscope digitized all 21 images of the blood smear cells. The blood cells were obtained from subjects with anaemia, and the algorithm has worked on 21 selected images with a size of 504 × 700 using ten samples in 40× magnification to detect the anaemia types with significant accuracy. Cells were thinly smeared onto glass microscope slides to enable individual cells types to be identified. Giemsa was employed in the blood preparations and stains; red blood cells in pink and platelets and white blood cells in magenta. The Nikon microscope specifications have been detailed;
the model is ECLIPSE Ci-E/Ci-L/Ci-S, CF160 infinity optical system, high-luminescent white LED Illuminator (Eco-illumination), automatic intensity reproduction function, and control image capture button. 6. Conclusions Microscopic image analysis of human blood smear cells has become an important tool for diagnosing healthy/unhealthy cells. Sickle, elliptocytosis, and microsite cells are the most important and common types of anaemia. This paper has presented a proposed algorithm that can detect and count the majority of healthy/unhealthy (sickle, elliptocytosis, microsites and unknown shapes) cells in a microscopic blood smear even if they are
40
H.A. Elsalamony / Micron 83 (2016) 32–41
Fig. 8. (a) The back propagation of the fourth NN unknown-shaped cells, (b) the best validation performance, and the confusion matrices in (c).
hidden or overlapped. The algorithm employs circular Hough transforms, morphological tools, and watershed segmentation to detect healthy and ill-shaped blood cells. The data of detected healthy/unhealthy cells, such as area, convex area, perimeter, eccentricity, solidity, and ratio, have been classified as input variables, whereas the functions TSC , TRC , MC and UC have been defined
as targets in four neural networks to analyse sickle, elliptocytosis, microsites and unknown-shaped cells, respectively. The experimental results have demonstrated high accuracy and success rates in analyzing healthy/unhealthy cells. Performance has been calculated by three statistical measures: classification accuracy, sensitivity, and specificity. This algorithm has been successful in
H.A. Elsalamony / Micron 83 (2016) 32–41
segmentation and classification processes for approximately 97.8% of all input cells, which may have contributed to improvements in the diagnosis process. The results indicate that an effectiveness of 100%, 98%, 100%, and 99.3% in neural networks has been achieved for various types of anaemia—sickle, elliptocytosis, microsites and unknown-shaped cells. Therefore, the proposed algorithm is very effective for diagnosing, counting, and detection processes for healthy/unhealthy red blood smear cells. References James, D., 13 July 2015. Microcytic anemia. Wikipedia web page, in: https://en. wikipedia.org/wiki/Microcytic anemia. 2015a. Sickle Cell Disease., Sickle Cell Disease Foundation of California webpage, USA in: http://www.scdfc.org/about-sickle-cell-disease.html. 2015b. What causes red blood cells to sickle, chemical and physical sciences practice passage questions. Khan Academy webpage., in https://www. khanacademy.org/test-prep/mcat/physical-sciences-practice. Biradar, N., Dewal, M.L., Rohit, M.K., 2015. Speckle noise reduction in b-mode echocardiographic images: a comparison. IETE Tech. Rev. 32, 435–453. Bishop, C.M., 2006. Pattern Recognition and Machine Learning. Springer, Singapore. Chaudhuri, B., Bhattacharya, U., 2000. Efficient training and improved performance of multilayer perceptron in pattern classification. Neurocomputing 34, 11–27. Das, D.K., Chakraborty, C., Mitra, B., Maiti, A.K., Ray, A.K., 2013. Quantitative microscopy approach for shape-based erythrocytes characterization in anaemia. J. Microsc. 249, 136–149. Deligiannidis, L., Arabnia, H., 2014. Emerging Trends in Image Processing, Computer Vision and Pattern Recognition. Morgan Kaufmann. Elsalamony, H.A., 2014. Sickle anemia and distorted blood cells detection using hough transform based on neural network and decision tree. In: Proceedings of the International Conference on Image Processing, Computer Vision, and Pattern Recognition (IPCV), The Steering Committee of The World Congress in Computer Science, Computer Engineering and Applied Computing (WorldComp).
41
Freeman, W.J., 1991. The physiology of perception. Sci. Am. 264, 78–85. Hirimutugoda, Y., Wijayarathna, G., 2010. Image analysis system for detection of red cell disorders using artificial neural networks. Sri Lanka J. Bio-Medic. Inf. 1, 35–42. Lam, P., 2015. Anemia: Causes, Symptoms and Treatments Medical News Today. MediLexicon, Intl., UK. Lee, H., Chen, Y.-P.P., 2014. Cell morphology based classification for red cells in blood smear images. Pattern Recognit. Lett. 49, 155–161. Mukherjee, R., 2014. Morphometric evaluation of preeclamptic placenta using light microscopic images. BioMed Res. Int. 9, 293690. Mushabe, M.C., Dendere, R., Douglas, T.S., 2013. Automated detection of malaria in Giemsa-stained thin blood smear. Engineering in Medicine and Biology Society (EMBC). 35th Annual International Conference of the IEEE, 3698–3701. Taherisadr, M., Nasirzonouzi, M., Baradaran, B., Mehdizade, A., Shiraz, I., 2013. New approch to red blood cell classification using morphological image processing. Shiraz E-Med. J., 14. Thirusittampalam, K., Hossain, M.J., Ghita, O., Whelan, P.F., 2013. A novel framework for cellular tracking and mitosis detection in dense phase contrast microscopy images. IEEE J. Biomed. Health Inform. 17, 642–653. Urrechaga, E., Hoffmann, J.J., Izquierdo, S., Escanero, J.F., 2015. Differential diagnosis of microcytic anemia: the role of microcytic and hypochromic erythrocytes. Int. J. Lab. Hematol. 37, 334–340. Weng, C.-H., Chen, J.-B., Wang, J., Wu, C.-C., Yu, Y., Lin, T.-H., 2011. Surgically curable non-iron deficiency microcytic anemia: Castleman’s disease. Oncol. Res. Treat. 34, 456–458. Xia, Y., Wu, X.-J., 2015. Adaptive ball particle filter and its application to visual tracking. IETE Tech. Rev. 32, 462–470. Yi, F., Moon, I., Lee, Y.H., 2015. Three-dimensional counting of morphologically normal human red blood cells via digital holographic microscopy. J. Biomed. Opt. 20, 016005.
Contents lists available at ScienceDirect
Micron journal homepage: www.elsevier.com/locate/micron
Healthy and unhealthy red blood cell detection in human blood smears using neural networks Hany A. Elsalamony a,b,∗ a b
Mathematics Department, Faculty of Science, Helwan University, Cairo, Egypt Computer Science & Information Department, Arts & Science College, Sattam University, Saudi Arabia
a r t i c l e
i n f o
Article history: Received 1 December 2015 Received in revised form 26 January 2016 Accepted 27 January 2016 Available online 1 February 2016 Keywords: Healthy/unhealthy RBC detection and counting Circular hough transforms Segmentation Neural network
a b s t r a c t One of the most common diseases that affect human red blood cells (RBCs) is anaemia. To diagnose anaemia, the following methods are typically employed: an identification process that is based on measuring the level of haemoglobin and the classification of RBCs based on a microscopic examination in blood smears. This paper presents a proposed algorithm for detecting and counting three types of anaemia-infected red blood cells in a microscopic coloured image using circular Hough transform and morphological tools. Anaemia cells include sickle, elliptocytosis, microsite cells and cells with unknown shapes. Additionally, the resulting data from the detection process have been analysed by a prevalent data analysis technique: the neural network. The experimental results for this model have demonstrated high accuracy for analysing healthy/unhealthy cells. This algorithm has achieved a maximum detection of approximately 97.8% of all cells in 21 microscopic images. Effectiveness rates of 100%, 98%, 100%, and 99.3% have been achieved using neural networks for sickle cells, elliptocytosis cells, microsite cells and cells with unknown shapes, respectively. © 2016 Published by Elsevier Ltd.
1. Introduction The general components of human blood are plasma, white blood cells (WBCs), red blood cells (RBCs) and platelets. RBCs comprise approximately 40% of blood volume. WBCs are smaller in volume but larger in size than RBCs. Plasma is the fluid component that contains melted salts and proteins. Platelet cells are similar particles but are smaller than WBCs and RBCs (Xia and Wu, 2015; Biradar et al., 2015; Deligiannidis and Arabnia, 2014). Anaemia is a type of red blood cell disorder that is usually caused by a lack of mineral iron in the blood. The human body needs iron to produce the iron-rich protein haemoglobin, which helps red blood cells carry oxygen from the lungs to the remainder of the body (Elsalamony, 2014; Das et al., 2013). This disease occurs when the blood has a lower than normal number of red blood cells (RBCs) or an insufficient amount of haemoglobin. RBCs are located inside the large bones of the body in the spongy marrow. The main function of marrow is to renew red blood cells, which continuously replaces old red blood cells. Normal RBCs die after they have lived in the bloodstream for 120 days. Their jobs include carrying oxygen and
∗ Correspondence to: Mathematics Department, Faculty of Science, Helwan University, Cairo, Egypt. E-mail addresses: h [email protected], [email protected] http://dx.doi.org/10.1016/j.micron.2016.01.008 0968-4328/© 2016 Published by Elsevier Ltd.
removing carbon dioxide (a waste product) from the body (Lam, 2015). RBCs are disc-shaped and can easily move through blood vessels. Elliptocytosis is a well-known type of anaemia. Historically, this disease was described in 1904 and recognized as a hereditary condition in 1932. The medical determination of hereditary elliptocytosis is difficult. The incidence of this disease ranges between three and five cases per 10,000 in the USA, whereas an estimated 60–150 cases per 10,000 of African and Mediterranean natives and 1500–2000 per 10,000 cases of Malayan natives have been documented (Lee and Chen, 2014). In sickle-cell anaemia, which is a serious disorder, the body creates a crescent shape of red blood cells. These sickle cells contain abnormal haemoglobin, which is referred to as sickle haemoglobin or haemoglobin S; it helps cells to develop a crescent shape. The absence of a polar amino acid encourages the noncovalent combination of haemoglobin in a low-oxygen environment, which distorts the red blood cells into a sickle shape and decreases their elasticity. Biochemically, the low-oxygen environment causes a chain of neighbouring haemoglobin molecules to hook together and block blood flow in the blood vessels of the limbs and organs, which become rigid and polymerized. Low blood flow can cause pain, organ damage, and increase the probability of disease. These cells fail to return to their normal shape when oxygen is restored and fail to deform as they pass through nar-
H.A. Elsalamony / Micron 83 (2016) 32–41
row vessels, which causes blockages in the capillaries ([2015b). Abnormal sickle cells usually die after approximately ten to 20 days. Bone marrow cannot renew red blood cells fast enough to replace dying red blood cells (Thirusittampalam et al., 2013). No cure exists for sickle cell anaemia. However, treatments can help to release pain and improve the complications of this disease. In addition, sickle-cell anaemia is common in people whose families derive from Mediterranean countries, Africa, South, or Central America, especially Panama, the Caribbean islands, Saudi Arabia, and India. In the United States, the disease derived from 70,000 to 100,000 people, primarily African Americans. The diagnosis of sickle cell anaemia is dependent on blood test analyses that can detect sickle cells (2015a). In the same context, microsites consist of small red blood cells. The most common causes of microcytic anaemia are iron deficiency and the thalassemia trait. The ability to distinguish microsites is very important for providing genetic counselling and preventing unnecessary and damaging iron therapy in thalassemia carriers. One of the simplest and most powerful discriminant functions is the ratio of microcytic cells to normal red blood cells (Urrechaga et al., 2015). The average size of a normal red blood cell is a mean corpuscular volume (MCV) in the range of 80–100 FL; smaller cells (90% total discrimination accuracy) perceived by FLDA, and PCA detected three critical principal components, which cumulatively elucidated 98.4% of the aggregate variance. In 2014, Lee and Chen (2014) introduced a neural network model with a classifier, which utilize the visual information obtained from the images of red blood cells images to determine if a red blood cell is standard or odd. They clustered the visual components into two essential parts—shape cluster groups and texture cluster groups—depending on the component properties. The input feature clusters were considered using parallel and course construction with input various data layers. Their trial results demonstrated imperative change and exactness in the proposed structure, which was distinguished from the single information layer classifier with element selection algorithms. In 2015, Yi et al. (2015) proposed a three-dimensional characterization strategy for consequently identifying the morpho-
34
H.A. Elsalamony / Micron 83 (2016) 32–41
Fig. 1. Difference in shape among normal RBCs.
logically ordinary RBCs in images of different human RBCs that were acquired by off-pivot computerized holographic microscopy (DHM). The 3D images of RBCs were recorded by DHM, and the stage images of various RBCs were subsequently reproduced by computational numerical calculation. They performed three runs of the typical RBC shapes—stomatocyte, discocyte, and echinocyte—for training and testing. The unusual RBC shapes were characterized as a fourth class. They obtained ten components, which were separated from every RBC, to fragment the reproduced stage images as a watershed transform, which include the projected surface region, average phase value, mean corpuscular haemoglobin, perimeter, mean corpuscular haemoglobin surface density, circularity, mean phase of centre part, sphericity coefficient, elongation, and pallor. They applied a principal component analysis algorithm to decrease the measurement number of variables and increase the Gaussian mixture densities by utilizing the projected data and the initial eight principal components. Therefore, the Gaussian mixtures were employed to plan the discriminant capacities based on Bayesian decision theory. Their experimental results demonstrated that the proposed strategy can achieve satisfactory results for ascertaining the rate of each average ordinary RBC shape in reconstructed stage images of numerous RBCs.
and (2)—to reduce errors in detection operation and ensure a final diagnosis of anaemia. The final stage of the proposed algorithm involves displaying the input image with all contoured healthy (green) and unhealthy (other colours) RBCs, in which all tested images are set to the same distance (focal length) from the microscope optics to control all measures. Via this operation, the data variables for each cell are measured, counted, and diagnosed for the majority of types in a smear. These measures serve as the input variables to train NN, whereas the output targets are measured based on the solidity (SC ), which is calculated by dividing the area by its convex area for all separate cells, as shown in Eq. (1), and the ratio (RC ), which is the ratio that is calculated by dividing the area of cells to the perimeter of cells, as shown in Eq. (2) SC =
areac (convexareac )
where areac is the area of each cell for all types of detected cells and convexareac represents the convex area of each cell for all types of detected healthy/unhealthy cells. RC =
3. The proposed algorithm The goal of this paper is to detect healthy and unhealthy blood cells in a coloured microscopic image of a human blood smear (Xia and Wu, 2015). The proposed algorithm is illustrated in Fig. 2. Fig. 2 presents an application of CHT with morphological functions on bright and dark intensity cells to detect and count healthy/unhealthy blood cells. CHT was employed to detect RBCs even if they were hidden, on image borders or overlapped. Then, watershed and morphological functions were employed to enhance and separate overlapped cells during the segmentation process. In CHT, many operations, such as cell polarity, which indicates whether circulating blood cells are brighter or darker than the background, and a computation method (two-stage), which calculates the accumulator array of CHT based on computing radial histograms, re apply estimated cell centres and the image information (Elsalamony). Another operation is the sensitivity factor, which is the sensor of the accumulator array in CHT. The detection process includes weak and partially hidden or overlapped cells. However, higher values of sensitivity increase the risk of false detection. The edge gradient threshold is the last operation that is performed; cells generally exhibit a darker interior (nuclei) that is surrounded by an outlying bright halo. The edge gradient threshold is very useful for determining edge pixels in images of both unhealthy and healthy blood cells based on their contrast, which are easily detected by establishing a lower threshold value. Fewer cells of weak edges are detected by increasing the value of the threshold (Deligiannidis and Arabnia, 2014). The NN has been applied to test and analyze the performance of the proposed algorithm for diagnosing and determining whether a patient has one or more types of anaemia. The classification is dependent on the data variables of the detected cells—area, convex area, perimeter, eccentricity, solidity, and ratio (refer to Eqs. (1)
(1)
4(areac )
(2)
2
(perimeterc )
The perimeterc represents the perimeter of each cell for all types of detected healthy/unhealthy cells. The target TSC is an array that consists of the binary numbers (1) and (0) for the conditions of solidity and eccentricity to classify sickle cell anaemia, as shown in Eq. (3).
TSC =
1 if SC < 0.72 or Eccentricity > 0.9 0
(3)
Otherwise
If any of the solidity values SC is less than 0.72 or the eccentricity is greater than 0.9, then the diagnosis is sickle cell anaemia. Otherwise, TSC assumes a value of zero with a diagnosis of healthy/unhealthy cells; however, the type anaemia is not sickle cell anaemia. The same conditions apply to target TRC , which is an array that is constructed in the conditions for ratio to classify elliptocytosis, as shown in Eq. (4).
TRC =
1 if 0.7 ≤ RC ≤ 0.9 0 Otherwise
(4)
when TRC is 1, the ratio RC ranges between 0.7 and 0.9, which indicates that the unhealthy cell belongs to elliptocytosis. When RC is less than 0.7 or greater than 0.9, then TRC is zero, which renders the decision that the cell is not derived from elliptocytosis and may be a healthy or other type of unhealthy cells. To detect microsites, Eq. (5) gives the condition that MC is 1 when the eccentricity exceeds 0.45 logical AND RC exceeds 0.9; otherwise, MC is zero.
MC =
1 if Eccentricity > 0.45 andRC > 0.9 0
Otherwise
(5)
H.A. Elsalamony / Micron 83 (2016) 32–41
35
Fig. 2. The main points of the proposed algorithm.
In Eq. (6), UC is an array that has been constructed in reverse conditions or conditions that complement all previous conditions in TSC , TRC and MC.
UC =
1 if Eccentricity > 0.9and SC ≥ 0.72 0
(6)
Otherwise
The target UC of unknown shapes is defined by 1 if the values of RC are less than 0.7, if logically and eccentricity are less than or equal to 0.9, and if SC is greater than or equal to 0.72. However, it assumes the value 0 when the value satisfies the conditions of TSC , TRC and MC ; then, the diagnosis is either healthy, sickle, elliptocytosis, or microsite. Table 1 lists values of variables according to healthy/unhealthy cells. The maximum and minimum values of the variables are listed according to the previously mentioned healthy/unhealthy cells. The minimum value of healthy cells is zero, which indicates that the cell is perfectly rounded or appears in a microscope in a complete position. Microsites are rounded with lower eccentricity values but smaller values than healthy cells. The back propagation neural network has trained and tested six variables as an input layer, ten neurons in the hidden layer and one neuron in the output layer. Three types of samples are applied: training samples, validation samples, and testing samples. The training samples represent 70% of all samples in the network during the training process in the input variables, and the network was modified according to the errors. Only 15% of the samples were employed in the validation to measure network generalization and to pause training when generalization stops improving. The remaining 15% of all samples of cells have been introduced as a testing sample, which has no effect on training and provides an independent measure of network performance during and after training. In addition, the mean square error (MSE), which is defined as the average squared difference between outputs and targets, was applied; lower values are better, and a value of zero indicates no error. Table 2 provides data on formed healthy/unhealthy cells. The range of all variables are similar, whereas only the output (TSC , TRC , MC , and UC ) variable has a binary value, such as 1, to exactly defined cells, and the remaining variables have a value of zero. The execution of every classification model has been assessed utilizing three factual measures: classification accuracy, sensitivity, and specificity. These benchmarks include the true positive (TP), true negative (TN), false positive (FP) and false negative (FN). A true positive choice occurs when the positive expectation of the classifier corresponds with a positive forecast of the previous segmentation. A true negative choice occurs when both the classifier and the division proposes the nonattendance of a positive expectation. A false positive choice occurs when the framework marks a healthy cell (positive expectation) as an unhealthy cell. A false negative choice occurs when the framework marks a negative (unhealthy) cell as a positive cell. In addition, the classification
accuracy has been characterized as a proportion of effectively ordered cells and is equivalent to all TP and TN that were isolated by the aggregate number of RBCs (N) (Bishop). Accuracy =
TP + TN N
(7)
The sensitivity alludes to the rate of correctly classified positive cells and is equivalent to the TP separated by the aggregate of TP and FN. Sensitivity =
TP TP + FN
(8)
Specificity refers to the rate of effectively arranged negative cells and is equivalent to the proportion of TN to the total TN and FP (Bishop). Specificity =
TN TN + FP
(9)
4. The experimental results The experimental results are displayed in two forms: the first form is the detection of healthy/unhealthy RBCs and the counting of these cells. The second form is the analysis and classification of the exported cells data to verify the agreement with the detection process using the most famous classification model in data analysis—the NN (Chaudhuri and Bhattacharya, 2000; Bishop, 2006). In the detection process, blood cells were obtained from subjects with anaemia, and the algorithm was employed using 21 selected images with a size of 504 × 700 for ten samples with 40× magnification to accurately detect the anaemia type—sickle, elliptocytosis, microsites, unknown shapes. Cells were thinly smeared onto glass magnifying lens slides to permit singular cell sorts to be recognized by utilizing the proposed calculation. Giemsa was utilized as a part of blood arrangements and stains: red blood cells (pink), and platelets and white blood cells (maroon).All cells (healthy/unhealthy) were detected using CHT, watershed, and morphological techniques for enhancement. In the same context, CHT was applied in the conditions of cell polarity to determine all dark and bright cells according to their intensity. Two-stage techniques were subsequently employed to compute the accumulator array of CHT. The sensitivity of this accumulator array in the proposed algorithm is 0.97 for brightness and 0.90 for dark cells, and an edge gradient threshold of 0.2 for the detection of fewer cells with weak edges. These conditions help CHT to detect the majority of the healthy RBCs that positioned singular, overlapped, and even attached to other unhealthy cells. Fig. 3(a) shows an example image of a human blood smear in, and some of the proposed algorithm steps are shown in (b) and (c). According to this image of the detection process, the number of healthy cells is 109 of 160, which is the total number of all detected cells (healthy/unhealthy). The remaining number of cells (51) may be considered as detection errors, unhealthy cells (sickle,
36
H.A. Elsalamony / Micron 83 (2016) 32–41
Table 1 Variables according to healthy/unhealthy cells. Variables
Healthy cells Unhealthy cells
Area
Convex area
Perimeter
Max
Min
Max
Min
Max
Min
Eccentricity Max
Min
Solidity Max
Min
Ratio Max
Min
797 869
539 158
821 1145
554 178
104.5685 150.3675
91.1127 51.1127
0.756632 0.97553
0 0.422419
0.97914 0.977337
0.940781 0.506494
0.929457 0.945956
0.815907 0.278963
Table 2 Blood cell variables. #
Variables
Type
Domain
1 2 3 4 5 6 7 8 9 10
Convex area Perimeter Eccentricity Areas Solidity Ratio TSC TRC MC UC
Ranged Ranged Ranged Ranged Ranged Ranged Binary Binary Binary Binary
178:1145 51.1127:150.3675 0:0.97553 158:869 0.506494:0.97914 0.278963:0.945956 (1 for sickle cell, 0 for other healthy/unhealthy cells) (1 for elliptocytosis, 0 for other healthy/unhealthy cells) (1 for microsites, 0 for other healthy/unhealthy cells) (1 for unknown shapes, 0 for other healthy/unhealthy cells)
Fig. 3. (a) The original image, (b) makes a small mark on every cell’s centroid, and (c) the final detected healthy cells.
elliptocytosis, microsites, and unknown), platelets, or even WBCs. Platelets and WBCs have been neglected (if they appeared in image) as the algorithm has only concentrated on healthy/unhealthy RBCs (Elsalamony). Additionally, the algorithm has determined 153 of the detected blood cells, without detection of unhealthy cells on the image borders, by approximately 97.8% according to the image in Fig. 3(a). Consequently, the next step is trying to classify these 51 cells into several types of unhealthy cells. All cells have strange shapes (crescent, elliptic, small RBCs, platelets, and unknown shapes); they are discovered and selected as unhealthy cells by the previous steps of healthy cell detection. Fig. 4 illustrates the detection of the majority of unhealthy cells without any cells in the image borders. In Fig. 4(a), a coloured segmented image for all unhealthy shapes with noise-like platelets. The sickle cells, elliptocytosis, microsites and unknown shapes have been detected in different contour colours, and the platelets have been neglected. The blue line contours represent the elliptocytosis cells, the cyan line represents microsites, the red line represents sickle cells (crescent or initiated to be crescent), whereas the yellow line represents unknown cell shapes. In the same context, of the 51 cells, the deformed cells include 15 sickle cells or initiated to be, 18 elliptocytosis cells, two of microsites and seven unknownshaped cells. The remaining seven unhealthy cells, which are not detected but may appear on the neglected image borders, include unknown detected cells that contain undefined or unknown composite shapes, or waste or unknown blood disease. In Fig. 4(b), the final detection of healthy cells are denoted by green lines and unhealthy cells are represented by a variety of coloured lines.
Experimentally, NN consists of six input variables and ten neurons in one hidden layer and one output layer. Three networks have trained with three targets—TSC , TRC and MC —for the detection of sickle, elliptocytosis and microsite cells; the fourth target has trained for unknown shapes with the target UC . The first NN is concerned with the detection of sickle cells, which is succeeded in 14 s after 29 iterations of 1000 as maximum iterations of the epoch, the performance 1.11e−07 , the gradient 7.67e−07 , and zero validation checks, as shown in Fig. 5(a). In Fig. 5(b), the best validation performance is shown as 3.6828e−09 at epoch 29, where the training is denoted by a blue line, the validation is denoted by a green line, and the test is denoted by a red line. The mean square error for training, validation, and testing is 1.10875e−07 , 3.6828e−09 , and 1.8058e−13 , respectively. The confusion matrices for the training, validation, and testing processes of the first NN are shown in (c). The predictions of the NN model are compared with the original classes of the target TSC to recognize the estimations of true positives, true negatives, false positives, and false negatives. These qualities have been processed to develop the confusion matrix, in which every cell contains the quantity of the cases that are arranged for the relating mix of craved and genuine classifier yields of 100%. An accuracy, sensitivity, and specificity of 100% is achieved. In the second NN model, the detection of elliptocytosis cells succeeded in 0 s after 56 iterations of 1000 as maximum iterations of the epoch, a performance of 0.00908, a gradient of 0.0643, and six validation checks, as shown in Fig. 6(a). In the same context, the best validation performance is shown as 0.029644 at epoch 50, where the training is denoted by a blue line, the validation is denoted by a green line, and the test is represented by the red line of (b). The mean square error of training, validation,
H.A. Elsalamony / Micron 83 (2016) 32–41
37
Fig. 4. (a) All contoured unhealthy cells, and (b) detection of all cells. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
Fig. 5. (a) The first NN Sickle cells, (b) the best validation performance, and the confusion matrices in (c). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
and testing are 1.02479e−02 , 2.9644e−02 , and 1.87565e−02 , respectively. The confusion matrices of the second NN with the target TRC are displayed in (c); 98% performance is achieved. Accordingly, accuracy, sensitivity, and specificity have achieved 99.1%, 96.4%, and 100%, respectively, due to the success of the training samples. They have achieved 95.7%, 75%, and 100% due to the success of the testing samples. The microsites NN has succeeded in 0 s after 23 iterations of 1000 as maximum iterations of the epoch, performance 6.72e−05 , gradi-
ent 0.000403, and six validation checks, as shown in Fig. 7(a). The best validation performance is 7.4256e−06 at epoch 17, as shown in Fig. 7(b). The mean square error for training, validation, and testing processes are 3.2056e−04 , 7.42563e−06 , and 1.62962e−03 , respectively. The confusion matrices of the third NN with the target MC are shown in (c); 100% performance is achieved. Accuracy, sensitivity, and specificity have achieved 100% due to the success of the training samples; they have achieved the same success of the testing samples.
38
H.A. Elsalamony / Micron 83 (2016) 32–41
Fig. 6. (a) The back propagation of the second NN elliptocytosis cells, (b) the best validation performance, and the confusion matrices in (c). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
The fourth NN is concerned with the detection of unknownshaped cells. It has succeeded, in 15 s, after 39 iterations of 1000 as maximum iterations of the epoch, performance of 0.000371, gradient of 0.00229, and six validation checks, as shown Fig. 8(a). Consequentially, the best validation performance is 0.0053972 at epoch 33, with the same coloured lines as shown in (b). The mean square error for training, validation, and testing are 1.60196e−03 , 5.39719e−03 , and 1.74046e−02 , respectively. The confusion matrices of the fourth NN with the target UC are shown in (c); 99.3%
performance is achieved. Accuracy, sensitivity, and specificity have achieved 100% due to the success of the training samples; however, they have achieved 95.7%, 100%, and 95.5% due to the success of the testing samples. The proposed algorithm is very effective in the diagnosing, counting, and detection of the healthy/unhealthy cells of a human blood smear.
H.A. Elsalamony / Micron 83 (2016) 32–41
39
Fig. 7. (a) The back propagation of third NN microsite cells, (b) the best validation performance, and the confusion matrices in (c).
5. Materials and devices Matlab-2013a was employed to construct the algorithm; a 32bit operating system with Windows 8, an Intel® CoreTM 2Duo CPU T5550 processor, 1.83 GHz and 2.50 GB RAM. The optical Nikon microscope digitized all 21 images of the blood smear cells. The blood cells were obtained from subjects with anaemia, and the algorithm has worked on 21 selected images with a size of 504 × 700 using ten samples in 40× magnification to detect the anaemia types with significant accuracy. Cells were thinly smeared onto glass microscope slides to enable individual cells types to be identified. Giemsa was employed in the blood preparations and stains; red blood cells in pink and platelets and white blood cells in magenta. The Nikon microscope specifications have been detailed;
the model is ECLIPSE Ci-E/Ci-L/Ci-S, CF160 infinity optical system, high-luminescent white LED Illuminator (Eco-illumination), automatic intensity reproduction function, and control image capture button. 6. Conclusions Microscopic image analysis of human blood smear cells has become an important tool for diagnosing healthy/unhealthy cells. Sickle, elliptocytosis, and microsite cells are the most important and common types of anaemia. This paper has presented a proposed algorithm that can detect and count the majority of healthy/unhealthy (sickle, elliptocytosis, microsites and unknown shapes) cells in a microscopic blood smear even if they are
40
H.A. Elsalamony / Micron 83 (2016) 32–41
Fig. 8. (a) The back propagation of the fourth NN unknown-shaped cells, (b) the best validation performance, and the confusion matrices in (c).
hidden or overlapped. The algorithm employs circular Hough transforms, morphological tools, and watershed segmentation to detect healthy and ill-shaped blood cells. The data of detected healthy/unhealthy cells, such as area, convex area, perimeter, eccentricity, solidity, and ratio, have been classified as input variables, whereas the functions TSC , TRC , MC and UC have been defined
as targets in four neural networks to analyse sickle, elliptocytosis, microsites and unknown-shaped cells, respectively. The experimental results have demonstrated high accuracy and success rates in analyzing healthy/unhealthy cells. Performance has been calculated by three statistical measures: classification accuracy, sensitivity, and specificity. This algorithm has been successful in
H.A. Elsalamony / Micron 83 (2016) 32–41
segmentation and classification processes for approximately 97.8% of all input cells, which may have contributed to improvements in the diagnosis process. The results indicate that an effectiveness of 100%, 98%, 100%, and 99.3% in neural networks has been achieved for various types of anaemia—sickle, elliptocytosis, microsites and unknown-shaped cells. Therefore, the proposed algorithm is very effective for diagnosing, counting, and detection processes for healthy/unhealthy red blood smear cells. References James, D., 13 July 2015. Microcytic anemia. Wikipedia web page, in: https://en. wikipedia.org/wiki/Microcytic anemia. 2015a. Sickle Cell Disease., Sickle Cell Disease Foundation of California webpage, USA in: http://www.scdfc.org/about-sickle-cell-disease.html. 2015b. What causes red blood cells to sickle, chemical and physical sciences practice passage questions. Khan Academy webpage., in https://www. khanacademy.org/test-prep/mcat/physical-sciences-practice. Biradar, N., Dewal, M.L., Rohit, M.K., 2015. Speckle noise reduction in b-mode echocardiographic images: a comparison. IETE Tech. Rev. 32, 435–453. Bishop, C.M., 2006. Pattern Recognition and Machine Learning. Springer, Singapore. Chaudhuri, B., Bhattacharya, U., 2000. Efficient training and improved performance of multilayer perceptron in pattern classification. Neurocomputing 34, 11–27. Das, D.K., Chakraborty, C., Mitra, B., Maiti, A.K., Ray, A.K., 2013. Quantitative microscopy approach for shape-based erythrocytes characterization in anaemia. J. Microsc. 249, 136–149. Deligiannidis, L., Arabnia, H., 2014. Emerging Trends in Image Processing, Computer Vision and Pattern Recognition. Morgan Kaufmann. Elsalamony, H.A., 2014. Sickle anemia and distorted blood cells detection using hough transform based on neural network and decision tree. In: Proceedings of the International Conference on Image Processing, Computer Vision, and Pattern Recognition (IPCV), The Steering Committee of The World Congress in Computer Science, Computer Engineering and Applied Computing (WorldComp).
41
Freeman, W.J., 1991. The physiology of perception. Sci. Am. 264, 78–85. Hirimutugoda, Y., Wijayarathna, G., 2010. Image analysis system for detection of red cell disorders using artificial neural networks. Sri Lanka J. Bio-Medic. Inf. 1, 35–42. Lam, P., 2015. Anemia: Causes, Symptoms and Treatments Medical News Today. MediLexicon, Intl., UK. Lee, H., Chen, Y.-P.P., 2014. Cell morphology based classification for red cells in blood smear images. Pattern Recognit. Lett. 49, 155–161. Mukherjee, R., 2014. Morphometric evaluation of preeclamptic placenta using light microscopic images. BioMed Res. Int. 9, 293690. Mushabe, M.C., Dendere, R., Douglas, T.S., 2013. Automated detection of malaria in Giemsa-stained thin blood smear. Engineering in Medicine and Biology Society (EMBC). 35th Annual International Conference of the IEEE, 3698–3701. Taherisadr, M., Nasirzonouzi, M., Baradaran, B., Mehdizade, A., Shiraz, I., 2013. New approch to red blood cell classification using morphological image processing. Shiraz E-Med. J., 14. Thirusittampalam, K., Hossain, M.J., Ghita, O., Whelan, P.F., 2013. A novel framework for cellular tracking and mitosis detection in dense phase contrast microscopy images. IEEE J. Biomed. Health Inform. 17, 642–653. Urrechaga, E., Hoffmann, J.J., Izquierdo, S., Escanero, J.F., 2015. Differential diagnosis of microcytic anemia: the role of microcytic and hypochromic erythrocytes. Int. J. Lab. Hematol. 37, 334–340. Weng, C.-H., Chen, J.-B., Wang, J., Wu, C.-C., Yu, Y., Lin, T.-H., 2011. Surgically curable non-iron deficiency microcytic anemia: Castleman’s disease. Oncol. Res. Treat. 34, 456–458. Xia, Y., Wu, X.-J., 2015. Adaptive ball particle filter and its application to visual tracking. IETE Tech. Rev. 32, 462–470. Yi, F., Moon, I., Lee, Y.H., 2015. Three-dimensional counting of morphologically normal human red blood cells via digital holographic microscopy. J. Biomed. Opt. 20, 016005.
