Image Segmentation of Acute Myeloid Leukemia Using Multi Otsu Thresholding

Acute Myeloid Leukemia (AML) can be identified by utilizing image processing. The stages used in the image processing process include preprocessing, segmentation, and feature extraction. The purpose of this segmentation is to separate objects in the image. One of the image segmentation methods is Multi Otsu Thresholding. This method has a way of working by looking for the threshold value dynamically. Previously there has been researched using the Static Otsu Thresholding method, but the results are still unsatisfactory because the threshold value obtained in the Static Thresholding method is that the threshold value obtained is static, meaning that the threshold value cannot adjust to the brightness level of each image that will be identified. With the background of the segmentation result dissatisfaction factor from the static threshold value, this study aims to carry out the segmentation process to obtain a dynamic threshold value, namely the Multi Otsu Thresholding Method. The Multi Otsu Thresholding method will be applied to segment AML images of M0 and M1 types, with the hope that the segmentation results obtained can be used for the feature extraction process better and are useful for the identification and classification process. Image processing methods used in this research are YCbCr color space, median filter, multi otsu thresholding, and morphological operations. The cell identification process utilizes the cell type classification process using the Naïve Bayes Classifier. The extracted characteristics were WBC (White Blood Cell) diameter, nucleus ratio, and nucleus roundness. Image data to test the process are 29 images of AML M0 and 30 images of AML M1. It was found that the segmentation results using the Multi Otsu Thresholding method can be used for the feature extraction process in AML M0 and AML M1 images and used for the identification process using the Naïve Bayes Classifier resulting in an accuracy of 83.81%, in this case, it has been compared with Static Otsu Thresholding which results with an accuracy of 75.35 %.


Introduction
Leukemia is a disease of the bone marrow that causes the production of immature cells or excessive blast of normal bone marrow [1]. AML is a malignant disease caused by immature cells which are often referred to as blast cells. AML is malignant because there are blast cells that divide rapidly. Although cells in AML do not divide as often as normal blast cells, AML cells will divide continuously like normal blast cells [2]. One way to assist the early diagnosis of AML is by utilizing image processing technology [3] [4] [5]). In image processing technology, image segmentation is a very crucial initial stage [6]. Image segmentation is the process of partitioning an image into several regions or objects [7]. One of the segmentation methods is the thresholding method, namely Static thresholding and Multi otsu thresholding [8]. Several studies on image segmentation have been conducted, including the segmentation process using the static thresholding method, but it has not  [9] [10]. The weakness of the static thresholding method is the difficulty in determining the threshold value used to separate objects in all existing images because each image has a different color intensity [9]. Another reason is that the different contrast in each image causes the segmentation results to be less accurate [4]. No method can determine with certainty the optimum threshold value for an image because the optimum threshold depends on the object of each image [10]. However, there is a method that can separate objects in the image quite well, namely the Multi Otsu Thresholding method which can separate leucocytes and produce nucleus and cytoplasmic regions accurately and can detect overlapping image objects because the threshold value selection is done dynamically based on a histogram. image [11] [12]. Based on the background that has been conveyed, this research aims to perform segmentation using Multi Otsu Thresholding. The difference with previous research is that in addition to the different data used, the previous research only focused on the segmentation of the image object [11] [12] whereas this research after the segmentation process was carried out, the segmentation results were used for the feature extraction process and continued for the classification process of cell types based on the characteristics that have been extracted for the identification process. The data used are images of acute leukemia types AML M0 and AML M1, then feature extraction is based on the diameter of WBC (White Blood Cell), nucleus ratio, and nucleus roundness. Furthermore, the feature extraction results will be used for the identification process whether the image is an image detected by AML M0 or AML M1 with the Naïve Bayes Classifier. Then based on the accuracy of the identification results, it will be evaluated how far the Otsu Thresholding Segmentation Method is. This method can segment cells on AML M0 and AML M1 images to start the process before the identification process.

Acute Myeloid Leukemia
Acute Myeloid Leukemia is a malignant, clonal disease that involves the proliferation of blast cells in the bone marrow, blood, or other tissues [1]. The blast cell types to be identified in this AML case are myeloblast, promyelocyte, and myelocyte [13]. AML M0 is characterized by a lack of clear myeloid differentiation on histological examination. In AML M0 there are almost no mature myeloid cells. The dominant blast cells are myeloblasts without granules or Auer rods. The nucleoli of blast cells are flattened and sometimes lobulated or split and contain fine chromatin with several different nucleoli [14]. AML M1 is a type of AML with the dominant cells being myeloblast cells more than 90%, found in the spinal cord with fine chromatin and a prominent nucleus shape, which may contain azurophilic granules (bluish color) [14].

Thresholding
Thresholding is a popular segmentation technique for monochrome images. One example of the segmentation method is Static Thresholding and Multi Otsu thresholding. Segmentation with Static Thresholding is carried out to separate the object from the background by determining the threshold value T. If the intensity value at point (x, y) is more than the threshold value T then it is marked as an object, whereas if it is less than the threshold value it is called the background. The threshold function can be defined in Equation 1. Thresholding results can be in the form of a binary image or a grayscale image [8].
Multiple Thresholding aims to divide the image into several regions or levels by using several threshold values [8]. In Multi Otsu Thresholding, the Otsu method is used to search for threshold values. In principle, as in bi-level Otsu, Multi-level Otsu also looks for the threshold value by maximizing the between-class variance. However, at the multi-level thresholding cumulative sum, not only two, but three, or four, etc. are calculated. The equation for calculating between-class variance can be defined in Equation 2 [8].

Naïve Bayes Classifier
Naïve Bayes is a statistical classification algorithm that assumes that there is no dependence between attributes [15]. This method approach uses the Bayes Theorem, where probability determines the possible outcomes of the classification process. Bayes' Theorem is expressed in Equation 3 [15].
For classification with continuous data used the Gauss Density Equation [16]. Gauss density can be calculated using Equation 4 [16].
Mean or average expressed in Equation 5 [16].

Research Method
The research is divided into several stages, it can be shown in Figure 1 below : Step of Research Figure 1 represents the stages of this research. The data inputted are the AML M0 and AML M1 image data obtained from the RS. Moewardi and identified by experts. Data collection in this study was carried out by collecting images of bone marrow indicated AML M0 and AML M1 from the Clinical Pathology Installation of Dr. Moewardi. The process of image data collection was obtained using a digital microscope, on bone marrow preparations with Giemsa staining. The number of bone marrow images taken was 59. Images were taken from 10 patients using a digital microscope with a magnification of 1000 times. Interviews with experts were also conducted to find information regarding the characteristics of AML types M0 and M1 along with the morphology of white blood cells infected with leukemia. The expert in this study was Doctor M.I Diah from RSUD Dr. Moewardi.

Data Input
The cells that will be identified in this study are myeloblast cells, promyelocyte cells, and myelocyte cells shown in Figure 2. Figure 2 was obtained from AML M1 image data collection at RSUD Dr. Moewardi Surakarta. From AML M0 images, only myeloblast and promyelocyte cells are found, while in AML M1 images, myeloblast, promyelocyte, and myelocyte cells can be found. In the AML M0 image, the doctor cannot identify all the cells contained in the image because the test image is bad. Myeloblast, promyelocyte, and myelocyte cells can be identified if the nucleus contains nucleoli. If a cell does not show nucleoli, doctors are not willing to identify the cell. Therefore, in the AML M1 image, only a few cells can be identified by the doctor. An example of nucleoli can be seen in Figure 3 [9].

Preprocessing
Pre-processing at this stage is carried out to obtain the same color distribution in each image. Preprocessing is done by using the median filtering method and then converting the RGB color to YCbCr. The image from the pre-processing process is then segmented using Multi Otsu Thresholding.

Segmentation
The segmentation process is carried out to obtain the same color distribution to separate the nucleus and WBC from the background. The segmentation process using Multi Otsu Thresholding is carried out in two stages, namely to detect the nucleus and detect WBC. The threshold value obtained is used to separate the nucleus and WBC. Subsequently, a morphological closing operation was performed to obtain the shape of the nucleus and the shape of the WBC. The next stage is the feature extraction process.

Feature Extraction
The feature extraction process is carried out to get the characteristics that will be used to detect AMLindicated white blood cells. The characteristics of white blood cells that are looking for are WBC diameter, nucleus ratio, and nucleus roundness. Calculations to get the quantity value of the cell's characteristics, among others: 1. WBC diameter, WBC diameter can be calculated based on the area of the detected WBC area. The WBC area obtained from the WBC detection process is shown in Equation 7.

@ABC.D.E = 2<
G H &IJK * 2. Nucleus ratio, the ratio of the nucleus is the ratio between the area of the nucleus and the area of the WBC. The nuclear ratio can be calculated by   (9) After obtaining all the characteristics of each cell, the classification process is then carried out to identify the types of cells in the image. The algorithm used is the Naïve Bayes Classifier. The parameters used for the classification process are the diameter of the WBC, the ratio of the nucleus, and the roundness of the nucleus which is the characteristic or feature of each cell in the image. Based on these three characteristics, it is then used to classify myeloblast, promyelocyte, myelocyte, and undefined cells.

Classification
After obtaining all the characteristics of each cell, the classification process is then carried out to identify the types of cells in the image. The algorithm used is the Naïve Bayes Classifier. The parameters used for the classification process are the diameter of the WBC, the ratio of the nucleus, and the roundness of the nucleus which is the characteristic or feature of each cell in the image. Based on these three characteristics, it is then used to classify myeloblast, promyelocyte, myelocyte, and undefined cells. The results of the accuracy of each cell are calculated by representing the classification results into a confusion matrix, namely true positive (TP), false negative (FN), or false positive (FP). The Confusion matrix table can be seen in Table 1.  Table 1 shows that:  TP (True Positive): the amount of data that the expert identified was correct and the system predicted to be correct  TN (True Negative): the amount of data that the expert identified was wrong and the system predicted incorrectly  FP (False Positive): the amount of data that the expert identified was false and the system predicted to be correct  FN (False Positive): the amount of data that the expert identified was true and the system predicted to be wrong.
Data from Table 1 will be represented in terms of accuracy, precision, and recall.
(12) Precision can represent over-segmentation while recall can represent under-segmentation. The higher the precision value, the lower the possibility of over-segmentation, while the higher the recall value, the lower the possibility of under-segmentation. Over-segmentation is a condition in which segmentation produces too many cells, while under-segmentation is a condition where segmentation produces too few cells.

Result and Discussion
The AML M0 and AML M1 data steps were obtained from RS. Moewardi and identified by experts. An example of the data in this study is shown in Figure 4. The application of multi otsu thresholding begins with pre-processing using the median filter and YCbCr color conversion. The results of smoothing with the median filter are then converted to YCbCr. YCbCr conversion aims to obtain the maximum value of cr (chrominance red) because the dominant color of white blood cells is red. The results on the median filter of the AML M1 image in Figure 4.b are shown in Figure 5 (a) and the YCbCr conversion is shown in Figure 5 (b).
Then do the conversion from the RGB color image in the image resulting from the median filter to grayscale. The results can be shown in Figure 6 below. Multi otsu thresholding has been able to separate the nucleus quite well. Subsequently, a closing operation was carried out on the segmentation results with a multi otsu threshold, as seen in Figure 7. While the results of the WBC segmentation and closing WBC can be seen in Figure 8   Feature extraction of the segmentation results was carried out by finding the area and perimeter (circumference) of the nucleus and the area of the WBC. The area of the nucleus, the perimeter of the nucleus, and the area of the WBC obtained were then used to calculate the diameter of the WBC, the ratio of the nucleus, and the roundness of the nucleus. The results of the multi otsu thresholding feature extraction can be seen in Figure 9.
The result of feature extraction in Figure 9 will be classified into four classes, namely myeloblast, promyelocyte, myelocyte, and unidentified cells. The classification results can be shown in Table 2.   The results of the identification of AML M0 and AML M1 images based on the classification results of the feature extraction of the segmented cells using MultI Otsu Thresholding can be shown in Table  3.  of the Multi Otsu Thresholding method are greater than the Static Thresholding method, indicating that the Multi Otsu Thresholding method tends to over-segmentation. In this study, over-segmentation occurs, which means a condition where there are too many segmented backgrounds as objects. From the test results that have been done segmentation with the Multi Otsu Thresholding method gives better classification results because Multi Otsu Thresholding applies the Otsu method which utilizes an image histogram which can dynamically select or determine the threshold value which can change dynamically adjusting the intensity of the color distribution in the image.

Conclusion
Based on the results and discussion that has been described, it can be concluded that the segmentation using the Multi Otsu Thresholding method gives quite good results compared to when used for segmentation of AML M0 and AML M1 images, which results in an accuracy of 83.81% and has been compared with the identification results obtained from segmentation of Static Thresholding which results are lower with an accuracy of 75.35%.