Med Biol Eng Comput (2015) 53:215–226 DOI 10.1007/s11517-014-1223-1

ORIGINAL ARTICLE

Effective identification and localization of immature precursors in bone marrow biopsy Guitao Cao · Ling Li · Weiting Chen · Yehua Yu · Jun Shi · Guixu Zhang · Xuehua Liu 

Received: 29 March 2013 / Accepted: 23 October 2014 / Published online: 28 November 2014 © International Federation for Medical and Biological Engineering 2014

Abstract  Abnormal localization of immature precursors (ALIP) aggregating and clustering in bone marrow biopsy appears earlier than that of bone marrow smears in detection of the relapse of acute myelocytic leukemia (AML). But traditional manual ALIP recognition has many shortcomings such as prone to false alarms, neglect of distribution law before three immature precursor cells gathered, and qualitative analysis instead of quantitative one. So, it is very important to develop a novel automatic method to identify and localize immature precursor cells for computer-aided diagnosis, to disclose their patterns before ALIP with the development of AML. The contributions of this paper are as follows. (1) After preprocessing the image with Otsu method, we identify both precursor cells and trabecular bone by multiple morphological operations and thresholds. (2) We localize the precursors in different regions according to their distances with the nearest trabecular bone based on chamfer distance transform, followed by discussion for the presumptions and limitations of our method. The accuracy of recognition and localization is

G. Cao · L. Li · W. Chen · X. Liu  Software Engineering Institute, East China Normal University, 3663 North Zhongshan Road, Shanghai 200062, China e-mail: [email protected] Y. Yu · J. Shi (*)  Sixth People’s Hospital, Affiliated to Shanghai Jiao Tong University, 600 Yishan Road, Shanghai 200233, China e-mail: [email protected] G. Zhang (*)  School of Information Science and Technology, East China Normal University, 500 Dongchuan Road, Shanghai 200241, China e-mail: [email protected]

evaluated based on a comparison with visual evaluation by two blinded observers. Keywords  Image segmentation · Morphology · Distance transform · Bone marrow biopsy · Acute myelocytic leukemia (AML)

1 Introduction Leukemia is a threat to human beings as one of the ten high incidences of malignant tumors, and relapse is the major factor affecting the survival rate of patients with acute myeloid leukemia (AML). Wang et al. [31] found that the appearance of abnormal localization of immature precursors (ALIP) aggregating and clustering in bone marrow biopsy is earlier than that of bone marrow smears in detection of AML. Therefore, there is a great need to research on the early diagnosis of relapse based on bone marrow pathology. The concept of ALIP coined by Tricot et al. in [27] indicated 3–5 or more primitive and immature granulocyte clusters that can be detected in trabecular bone area prompting the onset or recurrence of AML. Conventional ALIP detection relies mainly on manual observation under the microscope, including counting the number of single or clustering immature precursor cells per square millimeter using net-shaped micrometer and detecting the location of cluster cells by visual zoning method. The area between different trabecular bones is roughly divided into three regions, with region 1 next to the bone marrow area, region 2 corresponding to the central area between trabecular bones, and region 3 including the other cells. When 3–5 or above clustered immature precursors are found in region 2 or 3, they are considered as ALIP.

13

216

However, there are several shortcomings of manual ALIP recognition. Firstly, the workload of the cell recognition by eyes is prone to false alarms. Secondly, the distribution law is often ignored before three immature precursor cells gathered. Finally, visual localization leads to qualitative instead of quantitative analysis. So, it is a great need to develop a novel automatic algorithm not only to save diagnosis time for doctors, but also to reveal the changing pattern of number and localization before the arising of ALIP with the development of AML. To meet this requirement, we propose a novel algorithm to identify and locate the immature precursors and trabecular bone based on multi-morphology and distance transformation, revealing the distribution of single and cluster precursors.

2 Materials and methods 2.1 Case selection The 33 cases of AML patients (2 with M1, 10 with M2, 8 with M3, 6 with M4, 5 with M5, 1 with M6, and 1 with M7) is comprised of 18 males and 15 females. The age varied from 18 to 75 years, with median 28. There are 34 healthy cases, 20 males and 14 females with median age of 32 (varied from 20 to 70) for comparison. Bone marrow was collected by aspiration/trephine biopsy from AML patients during the complete remission in Shanghai Sixth People’s Hospital or Shanghai Changzheng Hospital from 2003 to 2008. They are diagnosed in-line with FAB criteria. 2.2 Image acquisition Bone marrow biopsies, performed under local anesthesia, are obtained using the conventional technique with a needle B65201 from the posterior superior iliac spines, fixed in Bouin solution, undecalcified, and gradient dehydrated using ethanol, followed by plastic Hemapun 865 embedding. Serial sections of 3 µm thickness are cut and stained by HGF (haematoxylin-Giemsa-acid fuchsin). Pathological images of 1,360 × 1,024 pixels are photographed by OLYMPUS microscope imaging system under 400 magnifications. 2.3 Methods Image segmentation refers to distinguish image from different regions with special meaning. Several works have been conducted on the segmentation of various types of pathology images (e.g., lymphoma, breast, and prostate cancers). They can be categorized by their features for segmentation, such as color, texture [11], Fourier transform [5], graph

13

Med Biol Eng Comput (2015) 53:215–226

[32], and contours [34]. Many popular color-based segmentation algorithms have been applied to medical images, including graph-cut [2], normalized-cut [26], Markov random field [12], and partial differential equation and level set [29], etc. Another aspect is to classify the segmentation methods by unsupervised algorithms, clustering [24], [35], expectation–maximization [25], and watershed [3, 6, 28, 9], etc. They tended to work well only when the cells have uniform nuclear regions. If the nuclear region has significant variation in color and texture, some holes will appear in the cells. Some researchers used morphological operations to fill these gaps [18]. So far, the cell segmentation focuses on the microscope images from bone marrow smear instead of pathology [1, 33, 19, 20, 10]. There are big difference in morphology of cell images  between smear and biopsy. From Fig. 1, we can see there are erythrocytes, leukocytes, myelocytes (myeloblast or immature precursors, neutrophilic myelocyte, and metagranulocyte), and trabecular bone in the bone marrow pathological image. The dyeing materials and individual difference make the automatic image segmentation a very tough work. Figure 1 shows that the characteristics of precursors are larger cell nucleus than other cells and obvious entoblasts,  and that of bone trabeculae is a big area in succession. So, we propose a new framework for immature precursors identification and localization as follows: • Image preprocessing, including denoising and binarization by Otsu method. • Immature precursor recognition and trabecular bone segmentation with morphological operations and different thresholds. • Single and cluster precursor cells discrimination and labeling. • Distance computation between the precursors and the nearest trabecular bone by chamfer distance transform. • Precursors localization into different regions. 2.3.1 Otsu method Otsu method [13, 22] aims to find the best separation between the two types of the image (e.g., foreground and background) by an optimal threshold. Given an intensity image X with L gray levels. If the pixel number of graylevel i is Ni, the total number of pixels in the image is N = L−1 i=0 Ni, and the probability of the i-th level will be Pi = Ni/N. Select a threshold k and divide all pixels into two categories (i.e., gray level ranging from 0 to k belong to Ω1 and from k +1 to L − 1 belong to Ω2, respectively), the mean gray level of entire image, Ω1 and Ω2 are µT =

L−1

i=0 iPi , µ1 (k) =

k

i=0 iPi , and µ2 (k) = µT − µ1 (k),

Med Biol Eng Comput (2015) 53:215–226

217

Erythrocyte

Dyeing materials

Leucocyte

Mature myelocyte (neutrophilic/ Immature myelocyte

Trabecular bone

Fig. 1  Bone marrow pathological image with different cells, trabecular bone, and dyeing materials

respectively. parts of the image area kThe ratios of thetwo L−1 are ω1 = i=0 Pi and ω2 = i=k+1 Pi = 1 − ω1, respectively. Let µ1 = µ1 (k)/ω1 and µ2 = µ2 (k)/ω2, intra-class variance σk2 can be defined as

σk2 = ω1 (µ1 − µT )2 + ω2 (µ2 − µT )2

(1)

Let threshold k changes from 0 to L  − 1, we can calculate the variance σk2 and find the largest value. Its corresponding k is the optimal threshold for image binarization. 2.3.2 Morphological methods Mathematical morphology based on set theory is to measure the shapes in the image with certain structural elements [21]. The basic transformations, expansion and corrosion, can be expressed as  A ⊕ B = {a + b : a ∈ A, b ∈ B} = Ab (2) b∈B

AΘB = {a : a + b ∈ A, b ∈ B} =

 b∈B

Ab

(3)

where A is the image, B is the structural element, and Θ are the expansion and corrosion operators, respectively. Their compositions, opening and closure operators, will be

A ◦ B = (AΘB) ⊕ B A • B = (A ⊕ B)ΘB

(4) (5)

In our research, the structure element for morphology operation is chosen as a circle with the radius of three pixels.

2.3.3 Chamfer distance transform In order to localize the cells, we have to calculate the distance of a cell to all pixels of the trabecular bone. An ideal way is to extract the edge of the trabecular bone and describe it as a function. But it is impossible because of the irregularity of trabecular bone. Another feasible method, distance transform, was put forward by Rosenfield and Pfaltz in [23] and was widely used in image understanding, pattern recognition, and computer vision [4, 8, 14, 17, 30]. Given a binary image, through Euclidean distance transform, it can be converted into a gray-level image with each value indicating the distance from the cell to its nearest bone trabecular. But the determination of the exact Euclidean distance transform is computation consume, chamfer distance, which is based upon “small” (i.e., 5 by 5) neighborhoods, plays an important role and outperforms in terms of computation time [16]. The coefficients of the chamfer mask C(k, l) (k, l = −2, −1, 0, 1, 2) are approximations of the local Euclidian distances separating the pixels covered by the mask. The pixel under analysis locates at the center of the mask; the local distance value adds to the corresponding pixel; and then the center pixel value updates. After m-th iteration, the minimum is the chamfer distance: m vi,j =

m−1 min (vi+k,j+l + C(k, l)), k, l = −2, −1, 0, 1, 2

(i,j)∈mask

(6)

m is the distance value of image pixel (i, j) in mwhere vi,j th iteration, and C(k, l) is the coefficients of the chamfer mask. The mask spreads to the whole image and iterates until all the distance values no longer change.

13

218

Med Biol Eng Comput (2015) 53:215–226

3 Experiment

Step 4 R2 OR operation with R3 to be optimization image R4 Step 5 Invert image R4 with the pixels interchanging between 0 and 1 to be R5 (see Fig. 3c)

3.1 Image preprocessing based on Otsu method The procedure of image preprocessing is as follows: Step 1  Change the bone marrow pathological image R0 (see Fig. 3a) into intensity image R1 by

Gray = 0.299 × R + 0.587 × G + 0.114 × B

(7)

where Gray indicates the gray level of the image, and R, G, and B are the red, green, and blue components of the color image, respectively Step 2 Using Otsu method for R1 to be binary image R2 (see Fig. 3b) Step 3 Denoise image R1 by median filter [7, 15] with a mask of 3 × 3 pixels, and then using Otsu as step 2 (R3)

Fig. 2  Flowchart of the cell recognition and labeling

The main processed images can be seen in Fig. 3. Figure  3a is a bone marrow pathological image R0; Fig. 3b shows R2 after step 2; and Fig. 3c stands for the preprocessed image R5 after step 3, 4, and 5.

3.2 Immature precursor cell recognition and labeling The flowchart of immature precursor cells recognition algorithm using morphology and threshold can be expressed as Fig. 2. The steps are as follows, and the experimental results are shown in Fig. 3d–i:

Preprocessed image

Step1: Closure operation to fill in the holes

Step2: Opening operation

Step4: Opening operation

Step3: Closure operation

Step5: Closure operation

Step3: OR operation

Step5: OR operation

Step6: Subtraction

Step7: Cell recognition

Step8: Precursors labeling

Step9: Cluster cells determination

13

Med Biol Eng Comput (2015) 53:215–226

219

Fig. 3  Examples of immature precursor cell preprocessing and recognition results: a Bone marrow pathological image R0, b binary image R2 by Otsu, c the preprocessed image R5, d image R7 after step 1–3, e image R8 after step 4–5, f subtracted image (step 6),

g cells recognition by area threshold (step 7), h cells edge extraction marked by blue circles (step 8), and i the precursors labeled with black dot, and cluster cells circled in black and their centers marked with green dot (step 9)

Step 1 Closure operation to fill in the holes on preprocessed image R5 to eliminate internal structures (R6) Step 2 Opening operation so as to split the adhesion regions Step 3 Closure operation and then OR image R6 to be image R7 (see Fig. 3d) Step 4 Opening operation to preprocessed image R5 Step 5 Closure operation and then OR operation with the image R6 to be image R8 (see Fig. 3e) Step 6 Subtract image R8 from the image R7 to remove the similar regions of these two images (see Fig.  3f). Because precursors have clear nucleolus (internal structures), the regions of other cells are

homogeneous. So, the regions of immature precursor can remain, and other cells will be lost after subtraction Step 7 Recognize the cells by threshold (see Fig. 3g). In our experiments, we can find that the size of immature precursor cells usually range from 200 pixels to 800 pixels Step 8  Label the immature precursor cells by their center coordinate and save the information of their edges (see Fig. 3h) Step 9  Determine the cluster cells. The distances between each cell center are calculated and compared to three times of the mean radius of all precursor cells. If the distance is less than the

13

220

Med Biol Eng Comput (2015) 53:215–226

Table 1  Detection results of the immature precursor cells recognition Sensitivity (%)

Specificity (%)

Accuracy (%)

(see Fig. 4c). Compute the area of enclosed regions, and mark the trabecular bones (see Fig. 4d).

90.4

96.8

94.7

3.4 Cells localization based on distance transformation The main steps of cell localization algorithm are as follows:

threshold, we label related cells as clusters. As shown in Fig. 3i, the precursors are labeled with black dot, and cluster cells are circled in black with their center marked with green dot

1. Chamfer distance transform for the segmentation image of bone trabecular.

Compared with current gold standard (visual evaluation of the marrow by two blinded observers), the sensitivity, specificity, and accuracy of the immature precursor cells recognition can be seen from Table 1.

We can  operate chamfer distance transform on the binary image of bone trabecular. Taking into account of the trabecular bone area as features and other pixel as background, the distance values of each pixel in the trabecular bone area are 0s; and the farther from the edge of trabecular bone, the greater the distance value of the background pixel will be.

3.3 Trabecular bone segmentation

2. Precursors localization based on their coordinates.

Due to the trabecular bone areas with larger and more homogeneous than the other parts, we segment them with a larger area threshold (i.e., 6,000 pixels). For a pathology image (see Fig. 4a), fill in the holes on its preprocessed image (see Fig. 4b) as the same as the cell recognition step 1–3 to be closed regions and without internal structures

After chamfer distance transformation, there exists a matrix with each value representing the minimum distance of the pixel from the trabecular bone. If there are varied trabeculae in a image, each background pixel will have more than one distance values. We usually choose the smaller one to represent the distance from the nearer trabeculae.

Fig. 4  Example of trabeculae bone segmentation process: a bone marrow pathological image, b Image after preprocessing, c filling in closed regions to eliminate the internal structure, and d the area of regions larger than 6,000 pixels are marked as trabecular bone

13

Med Biol Eng Comput (2015) 53:215–226

221

According to the coordinates of the detected cells, we can get their corresponding values from this matrix, that is, the distance from cells to the nearest trabecular bone. 3. Region division in accordance with the distance distribution. In our research, 1 pixel in the image is equal to the actual distance 0.88 μm. Compared with traditional visual

method, the surrounding area around the trabecular bone is divided into four regions, 0–200, 200–400, 400–600, and 600 μm–∞. One of the experimental images is shown in Fig. 5, with 14 single cells marked in red and 4 cluster cells marked in green, and trabecular bone in blue. There are two 0–200 μm regions corresponding to two trabecular bones. Another cells distribution image can be seen in Fig. 6 with three different trabeculae.

Fig. 5  Cells localization and region division: The area around trabecular bone (marked in blue) is divided into four regions (i.e., 0–200, 200–400, 400–600, and 600 μm–∞) with single precursors are marked in red, cluster in green, and trabecular bone in blue

Region 4 600μm

Region 3 400μm 600μm

Region 1 0 200μm

Region 1

Region 2

0 200μm

200μm 400μm

Fig. 6  Cells localization and region division: The area around trabecular bone (marked in blue) is divided into four regions (i.e., 0–200, 200–400, 400–600, and 600 μm–∞) with single precursors are marked in red, cluster in green, and trabecular bone in blue

Region 4 600μm Region 3

Region 1 0 200μm

Region 2

400μm 600μm

200μm 400μm

Region 1 0 200μm

Region 1 Region 3

0 200μm

400μm 600μm

13

222

4 Discussion Our method runs in windows XP/MATLAB7.0.1 within 10 s for the image of 1,360 × 1,024 pixels. Experiment results indicate that it can recognize and localize the immature precursor cells and trabeculae, with sensitivity of 90.4 %, specificity of 96.8 %, and accuracy of 94.7 % for the precursors. After we calculate the distance between cells and the trabecular bone, we can distribute the cells in different regions and analyze the numbers in particular one. It is more accurate than traditional visual zoning method. There are several parameters affecting the identification accuracy, such as thresholds for binary image using Otsu and area filtration as well. 1. Wight selection for image thresholding In our research, the threshold for image binarization is not derived directly from Otsu, but multiplied by a weight

Fig. 7  Weight selection results: horizontal axis is the weight H, the vertical axis the detection accuracy, and eight different color lines correspond to eight different images

Fig. 8  Selection of different maximal and minimal thresholds for area filtration, a maximal threshold experiment and b minimal threshold experiment: Abscissa indicates the maximal threshold in pixels of

13

Med Biol Eng Comput (2015) 53:215–226

H, ensuring the edge connectivity of the precursor cells or bone trabecula. Let H be 1, 1.01, 1.02, 1.025, 1.03, 1.035, and 1.05, we compare the detection accuracy by precursor cells, respectively. Figure 7 shows the correct detection rate by eight typical marrow pathology images, with vertical axis representing the detection accuracy and horizontal axis different H. We can see that after four attempts (H  = 1, 1.01, 1.02, and 1.025) at pushing the detection accuracy higher, the trend reverses and the accuracy heads lower. Maximum value appears when H is 1.025, which is what drove the interest in applying H of 1.025 for our experiments.

2. Thresholds for area filtration (A) Precursor area filtration In case of false detection for precursor cells because of having inner structures or vulnerable external influences, area filtration will play an important role. Compare the correct detection rates when the maximal area varies from 100, 200 to 2,000 pixels as shown in Fig. 8a, where abscissa indicates the maximal area in pixels; the vertical axis is the correct detection rate; and seven different color lines are typical seven bone marrow pathology images. It reveals that the detection rates increase rapidly when the area threshold are 100 and 200 pixels and decline significantly (i.e., line 6 and 7) when greater than 800 pixels. So we select 800 pixels as the maximal threshold in our research. Given fixed maximal area threshold of 800 pixels, we test their minimal threshold ranging from 0 to 150 as shown in Fig. 8b. When the area threshold is 95 pixels, detection rates reach the peak value. So we take 95 pixels as the minimal threshold in this paper.

area filtration; ordinate represents the correct detection rate; and each line corresponds to a typical bone marrow pathology images

Med Biol Eng Comput (2015) 53:215–226

223

(B) Area filtration for trabecular bone The procedure of area filtration for trabecular bone is the same as that for precursors with thresholds between 6,000 and 9,000 pixels. Figure 9 shows the examples by different area threshold TH for four trabecular bone images. It shows

that if the threshold is too large (i.e., 15,000 pixels), there will be missed trabeculae as shown in (b) and (e). If it is too small (i.e., 5,000 pixels), there exists false detected trabecular as shown in (h) and (k). When it is 6,000 pixels, we can get accurately recognition of the trabecular as shown in (c), (f), (i), and (l).

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

(i)

(j)

(k)

(l)

Fig. 9  Bone trabeculae detection by different threshold: Left original images, middle, and right detection results by different threshold such as TH = 15,000, 5,000, and 6,000 pixels, respectively, a original image 1, b TH = 15,000 pixels, c TH = 6,000 pixels, d origi-

nal image 2, e TH = 15,000 pixels, f TH = 6,000 pixels, g original image 3, h TH = 5,000 pixels, i TH = 6,000 pixels, j original image 4, k TH = 5,000 pixels, and l TH = 6,000 pixels

13

224

Med Biol Eng Comput (2015) 53:215–226

Fig. 10  Cells false detection because of noise and impurity, a experimental results and b visual evaluation by two blinded observers

noise

impurity

.

(a)

(b)

Missed detection

False detection

(a)

(b)

Fig. 11  Missed and false detected cells compared to visual evaluation by two blinded observers, a experimental results and b visual evaluation

3. Reasons for missed or false detection Detection accuracy decrease is derived from the impurity and the noise in the process of staining and imaging. In Fig. 10a, there are two false detections because of the impurity or noise with single precursors marked in red and clusters in green. Figure 10b is the visual evaluation by two blinded observers with immature precursor cells circled in black. Because of the non-uniform staining, the varying contrast between the nuclei/trabeculae and extracellular/ non-trabeculae regions across different sections, pathological images usually exhibit non-consistent colors, causing missed and false detection (see Figs. 11 and 12). In Fig. 11a, there is one missed and one false detected cell by

13

comparing with Fig. 11b, which is visually evaluated by two blinded observers with precursors that are circled in black. Figure  12a–c show three original images, (d)–(f) are their binary images, and (g)–(i) are their detection results with trabecular bone in blue, respectively. From Fig. 12d, e, g, and h, we can find that the trabecula can be accurately identified. But there is no trabecular area emerged in (f) and (i) because the trabecular bone regions share more similar characteristics with their surroundings. In fact, it is very difficult to discriminate erythroblasts with conspicuous nucleolus which often occurred in patients with AML and in CR, with real immature myeloblasts. In our experiments, we need to identify them by immunostaining, and our methods conform to the artificial evaluation.

Med Biol Eng Comput (2015) 53:215–226

225

(a)

(b)

(c)

(d)

(e)

(f)

undetected trabecular bone

(g)

(h)

(i)

Fig. 12  Trabeculae detection: a–c are three original images; d–f are the binary images of detected trabeculae; and g and h are the detected trabeculae marked in green, and there is no trabeculae detected in (i)

Acknowledgments  This work was supported by Natural Science Foundation of China (No. 61340036,81170507 and 81101119) and National Key Basic Research Program (No. 2011CB707104).

References 1. Al-Kofahi Y, Lassoued W, Lee W, Roysam B (2009) Improved automatic detection and segmentation of cell nuclei in histopathology images. IEEE Trans Biomed Eng 57(4):841–852 2. Boykov Y, Jolly MP (2001) Interactive graph cuts for optimal boundary and region segmentation of objects in ND images. Proc IEEE Int Conf Comput Vis 1:105–112 3. Cao GT, Zhong C, Li L, Dong J (2009) Detection of red blood cell in urine micrograph. In: The 3rd international conference on bioinformatics and biomedical engineering (ICBBE2009), Beijing, China, pp 1–4. doi:10.1109/ICBBE.2009.5162609 4. Corneliu T, Nedevschi S (2008) Real-time pedestrian classification exploiting 2D and 3D information. Intell Transp Syst IET 2:201–210

5. Fernandez DC, Bhargava R, Hewitt SM, Levin IW (2005) Infrared spectroscopic imaging for histopathologic recognition. Nat Biotechnol 23:469–474 6. Fu X, You H, Fu K (2012) A statistical approach to detect edges in SAR images based on square successive difference of averages. IEEE Trans Geosci Remote Sens Lett 9:1094–1098 7. Gopi Krishna S, Sreenivasulu Reddy T, Rajini GK (2012) Removal of high density salt and pepper noise through modified decision based unsymmetric trimmed median filter. Int J Eng Res Appl 2:090–094 8. Hajdu A, Hajdu L, Tijdeman R (2012) Approximation of the Euclidean distance by chamfer distances. Acta Cybern 20(3):399–417 9. Huang PW, Lai YH (2010) Effective segmentation and classification for HCC biopsy images. Pattern Recognit 43:1550–1563 10. Huang XW, Li H, Qiu YS, Xie SS (2006) Image processing technology in leukemia diagnosis. Laser Optoelectron Progress 43(10):42–46 11. Kong H, Gurcan M, Belkacem-Boussaid K (2011) Partitioning histopathological images: an integrated framework for supervised color-texture segmentation and cell splitting. IEEE Trans Med Imaging 30(9):1661–1677

13

226 12. Li SZ (2001) Markov random field modeling in computer vision. Springer, New York 13. Li L, Cao GT, Shi J, Wu H, Zhang XY (2010) Detecting immature precursor cells in pathological images of bone marrow based on morphology. 7th IEEE International Conference on Fuzzy Systems and Knowledge Discovery (FSKD2010), Engineering in Medicine and Biology Society, Beijing, China, pp 2190–2194. doi:10.1109/FSKD.2010.5569552 14. Marchand-Maillet S, Sharaiha YM (1999) Euclidean ordering via chamfer distance calculations. Comput Vis Image Underst 73:404–413 15. Morillas S, Gregori V, Sapena A (2011) Adaptive marginal median filter for colour images. Sensors 11:3205–3213 16. Mullikin JC (1992) The vector distance transform in two and three dimensions. CVGIP: Graph Model Image Process 54:526–535 17. Nedevschi S, Bota S, Tomiuc C (2009) Stereo-based pedestrian detection for collision-avoidance applications. IEEE Trans Intell Transp Syst 10:380–391 18. Nedzved A, Pitas I (2000) Morphological segmentation of histology cell images. IEEE Int Conf Pattern Recogn 1:500–503 19. Nilsson B, Heyden A (2005) Segmentation of complex cell clusters in microscopic images: application to bone marrow samples. Cytometry A 66(1):24–31. doi:10.1002/cyto.a.20153 20. Nipon TU (2005) White blood cell segmentation and classification in microscopic bone marrow images. In: Second International Conference on Fuzzy Systems and Knowledge Discovery (FSKD), pp 787–796 21. Ortiz F, Torres F (2004) Vectorial morphological reconstruction for brightness elimination in colour images. Real-Time Imaging 10:379–387 22. Otsu N (1979) A threshold selection method from gray-level histogram. IEEE Trans Syst Man Cybern Syst 9:62–66 23. Rosenfeld A, Pfaltz JL (1966) Sequential operations in digital picture processing. J Assoc Comput Mach 13:471–494 24. Sertel O, Kong J, Catalyurek U, Lozanski G, Saltz J, Gurcan M (2009) Histopathological image analysis using model-based

13

Med Biol Eng Comput (2015) 53:215–226 intermediate representation and color texture: follicular lymphoma grading. J Signal Process Syst 55(1):169–183 25. Sertel O, Lozanski G, Shana’ah A, Gurcan M (2010) Computeraided detection of centroblasts for follicular lymphoma grading using adaptive likelihood based cell segmentation. IEEE Trans Biomed Eng 57(10):2613–2616 26. Shi J, Malik J (2000) Normalized cuts and image segmentation. IEEE Trans Pattern Anal Mach Intell 22(8):888–905 27. Tricot G, De Wolf-Peeters C, Vlietinck R, Verwilghen RL (1984) Bone marrow histology in myelodysplastic syndrome prognostic value of abnormal localization of immature precursors in MDS. Br J Haematol 58(2):217–225 28. Vairalkar MK, Nimbhorkar SU (2012) Edge detection of images using sobel operator. Int J Emerg Technol Adv Eng 2:291–293 29. Vese LA, Chan TF (2002) A multiphase level set framework for image segmentation using the mumford and shah model. Int J Comput Vis 50(3):271–293 30. Wang J, Athitsos V, Sclaroff S, Betke M (2008) Detecting objects of variable shape structure with hidden state shape models. IEEE Trans Pattern Anal Mach Intell 30:477–492 31. Wang M, Shi J, Tao Y, Liu Y, Pu Q (2009) Correlation analysis of the characteristics of ALP with AML relapse. J Clin Hematol 22:345–346 32. Wen Q, Chang H, Parvin B (2009) A delaunay triangulation approach for segmenting clumps of nuclei. In: IEEE international symposium on biomedical imaging: from nano to macro, 2009 (ISBI09), pp 9–12. doi:10.1109/ISBI.2009.5192970 33. Yang DG, Dou WC, Cai SJ et al (2004) The application of HSI color model on the segmentation of karyocyte image. Comput Appl Softw 21(9):72–74 34. Yang L, Meer P, Foran D (2005) Unsupervised segmentation based on robust estimation and color active contour models. IEEE Trans Inf Technol Biomed 9(3):475–486 35. Zhou X, Liu KY, Bradley P, Perrimon N, Wong ST (2005) Towards automated cellular image segmentation for RNAI genome-wide screening. IEEE Conf Med Image Comput Comput Assist Interv 1:885–892