Patched Completed Local Binary Pattern is an Effective Method

for Neuroblastoma Histological Image Classification

Abstract on segmentation methods. First of all, different cells

in neuroblastoma histopathological images have vari-
Neuroblastoma is the most common extra cranial ations in illumination. Second, they have different
solid tumour in children. The histology of neuroblas- shapes within the same classification group which
toma has high intra-class variation, which misleads show the high intra-class variation of neuroblastoma.
existing computer-aided histological image classifica- In general, classification methods using segmentation
tion methods that use global features. To tackle this may fail to detect nuclei and the different cells in im-
problem, we propose a new Patched Completed Lo- ages because they segment different cells based on
cal Binary Pattern (PCLBP) method combining Sign illumination and shape. All existing feature-based
Binary Pattern (SBP) and Magnitude Binary Pat- classification methods used global features, extracted
tern (MBP) within local patches to build feature vec- from the whole image, which are sensitive to intra-
tors which are classified by k-Nearest Neighbor (k- class variations. To our best knowledge, there is no
NN) and Support Vector Machine (SVM) classifiers. feature-based method to address the intra-class vari-
The advantage of our method is extracting local fea- ation problem in neuroblastoma histological image
tures which are more robust to intra-class variation classification.
compared to global ones. We gathered a database of This paper proposes a local feature extraction
1043 histologic images of neuroblastic tumours clas- method for classification of neuroblastoma histologi-
sified into five subtypes. Our experiments show the cal images to tackle the intra-class variation problem.
proposed method improves the weighted average F- The contributions of this paper are:
measure by 1.89% and 0.81% with k-NN and SVM
classifiers, respectively. 1. We develop Patched Completed Local Binary
Pattern (PCLBP) based on Completed Local Bi-
Keywords: neuroblastic tumour, neuroblastoma, clas- nary Pattern (CLBP) (Guo et al. 2010) for clas-
sification, binary pattern, local patch, image analysis, sification of neuroblastoma histological images
computer-aided diagnosis (CAD) with intra-class variation.
2. We apply PCLBP on neuroblastoma histological
1 Introduction images with a complex histology to classify them
into five different categories.
Neuroblastoma is the most common extra cranial
solid tumour in children less than five years of age. 3. We evaluate our method by comparing with a
More than 15% of childhood cancer deaths are the state-of-the-art benchmark which shows the ef-
result of neuroblastic tumours (Park et al. 2008). Op- fectiveness of our method in the classification of
timal management of neuroblastic tumours depends neuroblastoma histological images.
on many factors, including histopathological classifi- The rest of this paper is as follows. Section 2
cation. Histological classification is performed by a presents related work. Section 3 describes the gath-
medical laboratory doctor who diagnoses tumours by ered dataset. Section 4 presents the proposed Patch
examining thin slices of tissue on a glass slide using an Completed Local Binary Pattern. Section 5 shows
optical microscope. Pathologists commonly use the the experimental results. Section 6 presents the dis-
Shimada system (Shimada et al. 1999) which iden- cussion and finally Section 7 concludes the work.
tifies six morphologic categories of neuroblastic tu-
mour. Computer image analysis of tumours has been
shown to improve diagnostic efficiency and consis- 2 Related Work
tency (Hipp et al. 2011), and identify previously un-
recognized image features that predict prognosis (Yu Most proposed techniques for classification of histo-
et al. 2016). Existing methods of computer-aided di- logical images rely on morphological features (Veta
agnosis (CAD) for classification of histological im- et al. 2014). However, variability of illumination
ages, thin slices of tissue mounted on a glass slide and appearance of different cells in the images makes
and viewed with a microscope, are divided into two classification based on segmentation more challeng-
categories: segmentation-based methods and feature- ing. Lee & Street (2003) proposed an automated
based methods. Segmentation-based methods rely on system for segmentation and classification of breast
morphological features such as symmetry. Feature- cancer’s nuclei. Mohammed et al. (2013) segmented
based methods try to extract mathematical features blood cell images using watershed technique to iden-
from the histological images and classify them with- tify chronic lymphocytic leukemia by extraction of
out segmentation (Boucheron 2008). nucleus and cytoplasm mask. Cosatto et al. (2008)
There are several factors that hinder the classi- classified breast cancer images using nuclei shape and
fication of neuroblastoma histological images based size. Nguyen et al. (2011) classified nuclei into nor-
mal and cancer based on appearance and identified
Figure 1: Neuroblastic tumour categories: (a) undifferentiated neuroblastoma, (b) poorly-differentiated neu-
roblastoma, (c) differentiating neuroblastoma, (d) ganglioneuroma, and (e) ganglioneuroblastoma

the cancer glands in prostate cancer. Sharma et al. logical images. Farjam et al. (2007) used texture fea-
(2015) segmented nuclei and classified with AdaBoost tures to identify the glandular elements within images
based on intensity and morphological features of nu- of prostate tissue. They applied k-means to cluster
clei. the image components. Spanhol et al. (2016) applied
Kong et al. (2009) classified neuroblastoma a completed modeling of the Local Binary Pattern
into three categories: undifferentiated, poorly- (LBP), based on three components extracted from
differentiated, and differentiating. They segmented the 8-neighborhood: center pixel, sign, and magni-
the images at each resolution level into cellular, tude. The center pixel is coded into a binary bit after
neuropil, and background elements and classified global thresholding. The difference signs and mag-
neuroblastoma histological images by integrating nitudes are coded in binary format so that they can
classifiers such as Linear Discriminative Analysis be combined to form the final Completed Local Bi-
(LDA) (Lehmann 1998), Support Vector Machine nary Pattern histograms. Spanhol et al. (2016) used
(SVM) (Cortes & Vapnik 1995), and k-Nearest Neigh- the combination of Completed Local Binary Patterns
bor (k-NN) (Cover & Hart 1967). Tafavogh et al. with k-NN and SVM to classify a large dataset of
(2014) proposed a four stage algorithm to classify breast cancer histological images into benign and ma-
neuroblastoma tumour images into undifferentiated lignant classes. Classification of neuroblastoma histo-
and poorly-differentiated using the Otsu segmenta- logical images remains challenging due to the intra-
tion technique (Otsu 1979). The whole slide image class variation.
was partitioned. Intensity variation was reduced us-
ing an image enhancement technique, regions of inter-
est were segmented by thresholding techniques and 3 Dataset
histological images were classified using rules based There is a lack of large and publicly available im-
on the Shimada scheme. age datasets for analysis of neuroblastic tumours,
In some applications, feature extraction has been which significantly hinders development and valida-
used to classify histological images. All of the existing tion of methods. Therefore, we gathered a dataset
methods extracted global features which are sensitive of images from neuroblastic tumours. Tumour ac-
to intra-class variation. Tabesh et al. (2007) classified cess is compliant with local policy, national legisla-
prostate tissue into tumour and non-tumour based tion, and ethical mandates to use the human tissue
on color, texture, and morphologic features. Qureshi in research. All patient specific details were removed
et al. (2008) classified meningioma using wavelet and a de-identified dataset was used for this research.
package transform and Local Binary Patterns (Ojala The initial dataset consisted of images of tissue mi-
et al. 2002). Zhang et al. (2013) used a combination of croarrays (TMA) of neuroblastic tumours, scanned by
curvelet transform, gray level co-occurrence matrix, the Aperio ScanScope system. Each slide was com-
and the Completed Local Binary Patterns (CLBP) as posed of 20 to 40 1.2mm cores of neuroblastic tumour,
features to stratify breast cancer tumours from histo- stained with haematoxylin and eosin (H&E) and cut
 Figure 2: Quantitative actual size of tissue spots and cropped images

Table 1: Number of different categories of neuroblastic tumour cropped images

Category of Neuroblastic Number of Number of
Tumour Cropped Images Patients

poorly-differentiated 571 77
differentiating 187 12
undifferentiated 155 10
ganglioneuroma 84 18
ganglioneuroblastoma 46 8

Total 1043 125

at 3µm. In this method, the contrast between dif- 4 Patch Completed Local Binary Pattern
ferent cells which have different colors is increased. (PCLBP)
Staining with H&E allows observation of histological
structures. TMA images were in svs format with res- Before describing our PCLBP algorithm, we first de-
olution 0.2µm, images were viewed and extracted us- scribe the algorithm it is based on Completed Local
ing ImageScope software (ImageScope 2016). Tissue Binary Pattern (CLBP) is one of the latest variants
cores were classified by experts into five different cate- of Local Binary Pattern (LBP) (Ojala et al. 2002).
gories: poorly-differentiated, differentiating, undiffer- The LBP operator computes the distribution of bi-
entiated, ganglioneuroma, and ganglioneuroblastoma, nary patterns in the circular neighborhood character-
according to the Shimada classification system. Rep- ized by a radius R and a number of neighbors P . The
resentative images in the categories are shown in Fig- idea is to threshold neighboring pixels, compared to
ure 1. the central pixel to the P neighbors. If the intensity
Areas best representative of each category, and de- of a neighbor pixel is greater than or equal to that
void of artefacts, were selected from each tissue core of the central pixel the value 1 is assigned, other-
by an expert histopathologist. At 40× magnification, wise 0. Therefore, a binary pattern is obtained from
cropped image size was 300×300 pixels with real size the neighborhood. The LBP function at pixel p is
80 × 80µm which is approximately one third of the (from (Ojala et al. 2002))
area of an optical microscope high power field of view.
Figure 2 shows the quantitative actual size of tissue P −1
spots and cropped images.
X
LBP (f (X, Y )) = 2i · u(f (Xi , Yi )−f (X, Y ))
This size was chosen as a compromise between be-
i=0
ing large enough to capture diagnostic features of each (1)
category and small enough for computational cost.
Numbers of images in our dataset are given in Table 1.
It is much larger in terms of patients and images than where f (Xi , Yi ) and f (X, Y ) are grey levels of pixels
the datasets used by Tafavogh et al. (2014) and Kong (Xi , Yi ) and (X, Y ) and u(·) is the unit step function.
et al. (2009). Moreover, the intra-class variation of The CLBP is a completed modeling of LBP (Guo
neuroblastoma cells in the gathered dataset is very et al. 2010) which is based on three components ex-
high which means different cells in neuroblastoma his- tracted from the local region: center pixel, sign, and
tological images within the same patients in the same magnitude. The center pixel is coded by a binary
class have different shapes. An example of the high code after thresholding, with the threshold set as the
intra-class variation of differentiating neuroblastoma average grey level of the whole image. For computing
is shown in Figure 3. As can be seen, both of them the sign and magnitude, a neighborhood of radius R
are differentiating type but their cells have different and number of neighbors P is considered. Signs and
shapes. magnitudes are computed and coded by a specific op-
erator into the binary format so that they can be
combined to form the final CLBP histograms (Guo
 Figure 3: An example of high intra-class variation of differentiating neuroblastoma

Figure 4: The scheme of the proposed method

et al. 2010). where f (X, Y ) denotes the original image. X pq , Y pq

Our approach, Patched Completed Local Binary denote the coordinate of the (p,q)th patch.
Pattern (PCLBP), extends CLBP. The overall frame-
work consists of four stages as shown in Figure 4.
First, the images are partitioned into equal-sized
square patches. Second, Sign Binary Patterns (SBPs)
and Magnitude Binary Patterns (MBPs) are com-
puted within patches. Third, histograms of SBPs and
MBPs are computed and concatenated to build a fea-
ture vector for each patch. A feature vector for the
whole image is created by concatenating the feature
vectors of all patches. Finally, the input image is clas-
sified by comparing the related feature vector with the
feature vectors of all images in the gallery. Following
we describe the algorithm in detail.
Given an N × N pixel input image, we partition
it into W × W pixel non-overlapping patches. We
indicate all points in the patch with p and q indices,
ranging from 1 to N/W , as
   
X Y
p= + 1, q = +1 (2)
W W
Figure 5: Coordinates in the patched image. Here,
where 0 ≤ X ≤ N and 0 ≤ Y ≤ N are the coor- we assume W = N/2
dinates of the input neuroblastoma image and b·c is
the floor function. The (p,q)th patch in the input The local differences of the 8-neighborhood around
neuroblastoma image (see Figure 5) is defined as (X0pq , Y0pq ), see Figure 6, are computed as
fp,q (X pq , Y pq ) = f (W (p−1)+X pq , W (q−1)+Y pq )
(3)
 f 0 p,q,k (X0pq , Y0pq ) =
fp,q (Xkpq , Ykpq ) − fp,q (X0pq , Y0pq ); k = 1, . . . , 8 (4)

Figure 6: An 8-neighborhood around (X0pq , Y0pq )

The SBP of the (p,q)th patch,

SBPp,q (fp,q (X0pq , Y0pq )), is defined as the con-
catenation of 8 bits as
Figure 7: An example of computed Magnitude Binary
Pattern (MBP) and Sign Binary Pattern (SBP): (a)
SBPp,q (fp,q (X0pq , Y0pq )) = original image, (b) MBP, and (c) SBP. The optimal
u(f 0 p,q,1 (X0pq , Y0pq )), . . . , u(f 0 p,q,8 (X0pq , Y0pq ))

(5) patch size (W ) is 60 Pixels. Here, we assume W = 150
pixels for better visualization.
where u(x) is the unit step function:
 and DSBP in an arbitrary neuroblastoma image. For
1 if x≥0 each patch, we model the distribution of DSBP and
u(x) = 0 if x<0 (6)
DMBP using the histogram operator with 256 bins
as
and the MBP for the (p,q)th patch is defined as

HSBPp,q (fp,q (X, Y )) = H{DSBPp,q (fp,q (X, Y ))}

M BPp,q (fp,q (X0pq , Y0pq )) = (10)
{u(f 0 p,q,1 (X0pq , Y0pq ) − m), . . . ,
u(f 0 p,q,8 (X0pq , Y0pq ) − m)} (7) HM BPp,q (fp,q (X, Y )) = H{DM BPp,q (fp,q (X, Y ))}
(11)
where m is a threshold to be set as the average of the We concatenate the histograms of the SBP and the
absolute values of all derivatives in the neuroblastoma MBP for each patch to build a Local Histogram (LH)
image. for each patch
After computing the SBP and the MBP, they are
converted into decimal values as LHp,q (fp,q (x, y)) =
{HSBPp,q (fp,q (x, y)), HM BPp,q (fp,q (x, y))} (12)
DSBPp,q (fp,q (X0pq , Y0pq )) =
8
X Then, we concatenate the histograms of all patches
2l−1 SBPp,q
l
(fp,q (X0pq , Y0pq )) (8) to build the Patch Completed Local Binary Pattern
l=1
for the neuroblastoma image (as shown in Figure 8)
P CLBP = {LHp,q (fp,q (X, Y )) |p, q = 1, ..., N/W }
(13)
DM BPp,q (fp,q (X0pq , Y0pq )) = Finally, two algorithms are used to classify the ex-
X 8 tracted PCLBPs: k-NN (Cover & Hart 1967) and
2l−1 M BPp,q l
(fp,q (X0pq , Y0pq )) (9) SVM (Cortes & Vapnik 1995).
l=1
5 Experimental Results
where SBPp,ql
(fp,q (X0pq , Y0pq )) and
l pq pq
M BPp,q (fp,q (X0 , Y0 )) denotes the l-th bit of In this section, we evaluate the performance of
the proposed method for classification of neuroblas-
the SBP and the MBP, respectively. DSBP and toma histological images. Experiments are conducted
DMBP for each pixel in the neuroblastoma image on the collected neuroblastic tumour database.
are computed. Figure 7 shows an example of DMBP
Figure 8: Concatenation of patches’ histograms: (a) patched image, (b) histogram of patches, and (c) concate-
nation of patches’ histograms.

Figure 9: Accuracy of k-NN classifier versus patch width (W ) and k in parameter tuning of k-NN classifier

The database is divided randomly into two sub- accuracies using different kernels. As can be seen,
sets: parameter-tuning (211 images) and validation the best result is achieved using the RBF kernel, so
datasets (832 images). We select the optimum values we selected it for the next experiments. RBF param-
for free parameters using the training dataset and fix eter γ was empirically defined through experiments
them for the validation. Then, we evaluate the system with best value taking 1/256 (256 is the number of
using the validation dataset and selected parameters. different intensities in the images).

5.1 Parameter Tuning

Table 2: Average classification accuracy of SVM over
We divide the parameter-tuning dataset into train- neuroblastic tumour dataset using different kernel
ing (150 images) and testing (61 images) subsets. We functions
train the algorithm using the training set with differ-
ent parameter values, test using the testing set and Kernel Classification Accuracy (%)
compute the accuracies. To have a better estima-
tion of the accuracy, we repeat the above procedure Linear 62.22
multiple (10) times and compute the average over all
experiments. Polynomial 71.82
For k-NN, the free parameters are the width of the
patch (W ) and the k numbers of neighbours. Accu- Radial Basis Function 72.33
racy was computed for k ranging from 1 to 10 and
patch width W ∈ {10, 15, 30, 50, 60, 75, 100, 150, 300} Sigmoid 49.45
and the results are shown in Figure 9. Best accuracy
of 72.7% was found with W =60 and k=5. So we used
these values in the next experiments. 5.2 System Validation
For SVM, we used the C-SVC type (Boser et al.
1992) using LIBSVM tool (Chang & Lin 2011) and Here, we use the remaining 80% of the dataset which
tested different kernels: linear, polynomial, Radial is not seen in the parameter tuning phase. It is di-
Basis Function (RBF), and sigmoid. Table 2 shows vided into training (623 images) and validation (209
 Figure 10: Comparison between our algorithm (PCLBP) and Spanhol’s system (CLBP)

Table 3: Weighted average precision, recall, and F-measure obtained by our system and Spanhol’s system.

Our system (PCLBP)(%) Spanhol’s system (CLBP)(%)

Precision-kNN 70.49±3.37 68.22±3.23

Precision-SVM 75.59±3.15 74.1±2.35

Recall-kNN 71.02±2.87 69.53±2.81

Recall-SVM 76.35±3.41 76.25±2.23

F-measure-kNN 70.75±3.09 68.86±6.04

F-measure-SVM 75.96±3.27 75.15±2.28

images) sets. We train the algorithm using the train- ages to five different categories.
ing set (with the parameter values selected in Sec-
tion 5.1) and test using the validation set. We repeat 2. Neuroblastoma has a complex texture with a
this procedure multiple (10) times and report the av- great deal of complicated features compared to
erage accuracy. Algorithm performance for k-NN and other types of cancer such as breast cancer. It
SVM is reported with the average F-measure, recall, is the first time that neuroblastoma histological
and precision (Powers 2011) weighted by number of images are classified into five different categories
examples in each of the five classes. Distribution of using a feature extraction method.
the computed F-measures for the two classifiers and 3. The proposed method extract features within
feature extraction approaches over the ten trials is small patches which are not easily detected by
presented in Figure 10 and shows that SVM works human eyes.
better than k-NN. The t-test with P value = 0.03
and α=0.05 (significance level) shows that combina-
tion of our algorithm with SVM classifier significantly 7 Conclusion
improves the accuracy of classification in comparison
with k-NN classifier. Table 3 indicates that our al- We proposed a new Patched Completed Local Binary
gorithm obtains approximately 5% higher accuracy Pattern (PCLBP) to classify neuroblastic tumours
when it is combined with SVM classifier compared into five different categories using extracted feature
to the k-NN classifier. We also test the CLBP algo- vectors from histological images. The algorithm built
rithm (Spanhol et al. 2016) on the test images as a the feature vector by extraction of SBP and MBP
benchmark, again reporting the weighted average of within local patches. The advantage of the proposed
precision, recall, and F-measure. Table 3 reports the method is extraction of local features which are more
weighted average precision, recall, and F-measure of robust to intra-class variation compared to global fea-
our system and Spanhol’s system. The weighted av- ture extraction. The evaluation was conducted on a
erage precision, recall, and F-measure of our system gathered dataset with 1043 cropped images from sam-
are better than Spanhol’s system. ples of five different categories. We compare the re-
sults obtained by our system with the state-of-the-art.
Results indicate that the proposed method has im-
6 Discussion proved the average weighted F-measure for k-NN and
SVM by 1.89% and 0.81%, respectively, compared to
The proposed algorithm is a new feature extraction the benchmark.
method to classify neuroblastoma histological images
into five different groups. Although a large number
of methods have been proposed in the literature, our References
system has multiple advantages over these systems:
Boser, B., Guyon, I. & Vapnik, V. (1992), A Training
1. There is no feature based method to classify Algorithm for Optimal Margin Classifiers, in ‘Pro-
histological images into more than two cate- ceedings of the Fifth Annual Workshop on Compu-
gories. They were classified more straightforward tational Learning Theory’, pp. 144–152.
to binary classification. However, the proposed
method can classify neuroblastoma tumour im- Boucheron, L. (2008), Object-and Spatial-Level Quan-
titative Analysis of Multispectral Histopathology
 Images for Detection and Characterization of Can- Ojala, T., Pietikainen, M. & Maenpaa, T. (2002),
cer, Doctoral Dissertation in University of Califor- ‘Multiresolution Gray-Scale and Rotation Invariant
nia at Santa Barbara. Texture Classification with Local Binary Patterns’,
IEEE Transactions on Pattern Analysis and Ma-
Chang, C. & Lin, C. (2011), ‘LIBSVM: A Library for chine Intelligence 24(7), 971–987.
Support Vector Machines’, ACM Transactions on
Intelligent Systems and Technology (TIST) 2(3), 1– Otsu, N. (1979), ‘A Threshold Selection Method
27. Software available at http://www.csie.ntu. from Gray-Level Histograms’, IEEE Transactions
edu.tw/~cjlin/libsvm. on Systems, Man, and Cybernetics 9(1), 62–66.
Cortes, C. & Vapnik, V. (1995), ‘Support-Vector Net- Park, J., Caron, H. & Eggert, A. (2008), ‘Neurob-
works’, Machine Learning 20(3), 273–297. lastoma: Biology, Prognosis, and Treatment’, Pe-
diatric Clinics of North America 55(1), 97–120.
Cosatto, E., Miller, M., Graf, H. & Meyer, J. (2008),
Grading Nuclear Pleomorphism on Histological Mi- Powers, D. (2011), ‘Evaluation: from precision, re-
crographs, in ‘International Conference on Pattern call and f-measure to roc, informedness, marked-
Recognition’, pp. 1–4. ness, and correlation’, Journal of Machine Learning
Technologies 2(1), 37–63.
Cover, T. & Hart, P. (1967), ‘Nearest Neighbor Pat-
tern Classification’, IEEE Transactions on Infor- Qureshi, H., Sertel, O., Rajpoot, N., Wilson, R. &
mation Theory 13(1), 21–27. Gurcan, M. (2008), Adaptive Discriminant Wavelet
Packet Transform and Local Binary Patterns for
Farjam, R., Soltanian-Zadeh, H., Jafari-Khouzani, Meningioma Subtype Classification, in ‘Proceed-
K. & Zoroofi, R. (2007), ‘An Image Analysis Ap- ings of the 11th International Conference on Med-
proach for Automatic Malignancy Determination of ical Image Computing and Computer-Assisted In-
Prostate Pathological Images’, Cytometry Part B: tervention, Part II’, pp. 196–204.
Clinical Cytometry 72(4), 227–240.
Sharma, H., Zerbe, N., Heim, D., Wienert, S.,
Guo, Z., Zhang, L. & Zhang, D. (2010), ‘A Com- Behrens, H., Hellwich, O. & Hufnagl, P. (2015), A
pleted Modeling of Local Binary Pattern Operator Multi-Resolution Approach for Combining Visual
for Texture Classification’, IEEE Transactions on Information using Nuclei Segmentation and Clas-
Image Processing 19(6), 1657–1663. sification in Histopathological Images, in ‘10th In-
Hipp, J., Flotte, T., Monaco, J., Cheng, J., Mad- ternational Conference on Computer Vision Theory
abhushi, A., Yagi, Y., Rodriguez-Canales, J., and Applications’, pp. 37–46.
Emmert-Buck, M., Dugan, M., Hewitt, S., Toner, Shimada, H., Ambros, I., Dehner, L., Hata, J., Joshi,
M., Tompkins, R., Lucas, D., Gilbertson, J. V., Roald, B., Stram, D., Gerbing, R., Lukens,
& Balis, U. (2011), ‘Computer-Aided Diagnos- J., Matthay, K. & Castleberry, R. (1999), ‘The
tic Tools Aim to Empower Rather than Replace International Neuroblastoma Pathology Classifica-
Pathologists: Lessons Learned from Computational tion (the Shimada System)’, Cancer 86(2), 364–
Chess’, Journal of Pathology Informatics 2(1), 25– 372.
25.
Spanhol, F., Oliveira, L., Caroline, P. & Laurent, H.
ImageScope (2016). (2016), ‘A Dataset for Breast Cancer Histopatho-
URL: http://www.leicabiosystems.com/digital- logical Image Classification’, IEEE Transactions on
pathology/digital-pathology- Biomedical Engineering 63(7), 1455–1462.
management/imagescope/
Tabesh, A., Teverovskiy, M., Pang, H., Kumar, V.,
Kong, J., Sertel, O., Shimada, H., Boyer, K., Saltz, Verbel, D., Kotsianti, A. & Saidi, O. (2007), ‘Mul-
J. & Gurcan, M. (2009), ‘Computer-Aided Evalu- tifeature Prostate Cancer Diagnosis and Gleason
ation of Neuroblastoma on Whole-Slide Histology Grading of Histological Images’, IEEE Transac-
Images: Classifying Grade of Neuroblastic Differ- tions on Medical Imaging 26(10), 1366–1378.
entiation’, Pattern Recognition 42(6), 1080–1092.
Tafavogh, S., Meng, Q., Catchpoole, D. & Kennedy,
Lee, K. & Street, W. (2003), ‘An Adaptive Resource- P. (2014), Automated quantitative and qualitative
Allocating Network for Automated Detection, Seg- analysis of the whole slide images of neuroblastoma
mentation, and Classification of Breast Cancer Nu- tumour for making a prognosis decision, in ‘Pro-
clei Topic Area: Image Processing and Recog- ceedings of the IASTED 11th International Con-
nition’, IEEE Transactions on Neural Networks ference on Biomedical Engineering’, pp. 244–251.
14(3), 680–687.
Veta, M., Pluim, J., van Diest, P. & Viergever, M.
Lehmann, E. (1998), Theory of Point Estimation, (2014), ‘Breast Cancer Histopathology Image Anal-
Springer. ysis: A Review’, IEEE Transactions Biomed Eng.
Mohammed, E., Mohamed, M., Naugler, C. & Far, B. 61(5), 1400–1411.
(2013), Chronic Lymphocytic Leukemia Cell Seg- Yu, K., Zhang, C., Berry, G., Altman, R., Re, C.,
mentation from Microscopic Blood Images using Rubin, D. & Snyder, M. (2016), ‘Predicting Non-
Watershed Algorithm and Optimal Thresholding, Small Cell Lung Cancer Prognosis by Fully Auto-
in ‘26th IEEE Canadian Conference on Electrical mated Microscopic Pathology Image Features’, Na-
and Computer Engineering’, pp. 1–5. ture Communications 7, 1–10.
Nguyen, K., Jain, A. & Sabata, B. (2011), ‘Prostate Zhang, Y., Zhang, B. & Lu, W. (2013), ‘Breast
Cancer Detection: Fusion of Cytological and Tex- Cancer Histological Image Classification with Mul-
tural Features’, Journal of Pathology Informatics tiple Features and Random Subspace Classifier
2:S3. Ensemble’, Studies in Computational Intelligence
450, 27–42.