Multimedia Tools and Applications (2020) 79:31049–31068

https://doi.org/10.1007/s11042-020-09640-9

Active contour modal based on density-oriented BIRCH

clustering method for medical image segmentation

Shoulin Yin 1 1 2
& Hang Li & Desheng Liu & Shahid Karim
3

Received: 13 November 2019 / Revised: 3 August 2020 / Accepted: 13 August 2020 /

Published online: 18 August 2020
# Springer Science+Business Media, LLC, part of Springer Nature 2020

Abstract
Currently, medical image segmentation has attracted more attention from researchers,
which can assist in medical diagnosis. However, in the process of traditional medical
image segmentation, it is sensitive to the initial contour and noise, which is difficult to
deal with the weak edge image, complex iterative process. In this paper, we propose a
new medical image segmentation method, which adopts density-oriented BIRCH (bal-
anced iterative reducing and clustering using hierarchies) clustering method to modify
active contour model and improve the robustness of noise. The BIRCH is a multi-stage
clustering method using clustering feature tree. The improved model can effectively deal
with the gray non-uniformity of real medical images. And we also introduce a new energy
function in active contour model to make the contour curve approach to the edge, and
finally stay at the edge of the image to complete the object segmentation. Experimental
results show that this new model can overcome the influence of complex background on
medical image segmentation and improve the speed and accuracy of medical segmenta-
tion results.

Keywords Medical image segmentation . Density-oriented BIRCH . Active contour model .

Energy function

* Hang Li
[email protected]

Shoulin Yin
[email protected]

Desheng Liu
[email protected]
Shahid Karim
[email protected]

Extended author information available on the last page of the article

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

31050 Multimedia Tools and Applications (2020) 79:31049–31068

1 Introduction

With the development of medical imaging technology, biomedical image segmentation is

becoming more and more important in clinical diagnosis and treatment [15]. Active contour
models are widely used in medical image segmentation due to the complex grayscale of
medical images and the variety of topological structures. In order to better deal with the
problems in the field of image segmentation, especially for medical image segmentation, the
level set method is introduced, which greatly improves the difficulty of complex image
processing [11, 33]. Nowadays, the level set method is widely used in image processing and
other fields, and the edge set, region level set image segmentation algorithms are proposed.
The boundary-based active contour model is mainly derived from image gradient information,
so it is sensitive to noise, initial contour position and gray level inhomogeneity. The region-
based active contour model uses regional information to reduce the sensitivity to noise,
effectively overcome the above deficiencies.
As for the non-uniformity of image grayscale, the LBF (Local Binary Fitting) model based
on Local area Fitting is proposed [19]. Although the image presents uneven gray level in the
global scope, the gray level changes slowly in the small neighborhood area. Therefore, the use
of local mean information in LBF model can better reduce the impact of image grayscale
inhomogeneity, and has been widely used in medical image segmentation. However, this
model is not ideal for image segmentation with strong noise and weak boundary. To solve the
above problem, Zhang [32] proposed a GAC model based on global information, which used
the pressure function based on global information to replace the boundary stop function in the
GAC model (geographical active contour). It was good at image segmentation with noise and
weak boundary. However, it was still based on the uniform distribution of the gray level of the
image, which was not ideal for the uneven gray level image. Aiming at the problems of
migration field, noise and weak boundary in images, Singh [20] introduced window function
and localized the symbolic pressure function, so as to better reflect the gray level change of
medical images and achieve better segmentation effect for images with migration field. The
study found that the local gray level information of most medical images presented a Gaussian
distribution. In response to this phenomenon, Bonini [3] further proposed a small neighbor-
hood probability active contour model, namely LGD model. The model fitted the distribution
of image’s small neighborhood well, and had good segmentation effect for the image with
offset field, but it was easy to be affected by the image background quality. So in this paper,
we propose a new medical image segment method to address the defects of the above methods.
At the same time, the clustering algorithm [12] in image processing has unique advantages
for fast coarse-grained image processing. In order to improve segmentation accuracy, com-
bined with density-oriented BIRCH clustering, a new algorithm model is formed. The
accuracy of medical image segmentation is effectively improved with proposed method, and
the image processing effect is verified experimentally. Motivated by the recent success of
depthwise separable active contour model, we also explore this operation and show improve-
ment in terms of both speed and accuracy by adapting the BIRCH method. Finally, we
demonstrate the effectiveness of the proposed model on medical data and attain the better test
set performance compared to the state-of-the-art segment methods.
In summary, our contributions are:

a. We propose a new medical image segmentation which employs density-oriented BIRCH

clustering method to modify active contour model.

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Multimedia Tools and Applications (2020) 79:31049–31068 31051

b. In our model, one can arbitrarily change the resolution or size of medical images.
c. The BIRCH is a multi-stage clustering method using clustering feature tree. The improved
model can effectively deal with the gray non-uniformity of real medical images.
d. We also introduce a new energy function in active contour model to make the contour
curve approach to the edge.
e. Our proposed model attains a new state-of-art performance on medical datasets. We also
provide detailed analysis of design choices and model variants.

The structure of our paper is organized as follows. In Section 2, we present the related works.
Then, the active contour model and cluster are presented in Sections 3. In section 4, we
detailed introduce the proposed segment method. Finally, Section 5 shows the experiments and
section 6 presents the conclusion, respectively.

2 Related works

The active contour model has been successfully applied into many fields such as object
detection, medical image processing, remote sensing, etc., It combines the attributes (such as
gray-scale and edges of the image) with the prior information (such as shape, colour) to form
energy functionals and finally obtain the complete expression of the segmented object. Koh [9]
proposed a gradient vector flow (GVF), which effectively improved the initial contour to be
close to the target boundary. But the convergence effect at the concave edge was not ideal.
Wang [24] proposed a geometric active contour model. This type of model was widely used. It
was based on the curve evolution theory and the level set method. It used the variational
method to minimize the energy functional, which solved the problem of “boundary leakage” to
a certain extent. And most proposed active contour models are based on level set methods and
have unique model advantage for complex medical images.
The above is an edge-based active contour model that is sensitive to noise. Mumford and
Shah proposed the M-S model, which was the earliest regional active contour model, but the
mathematical knowledge was deep and difficult to understand. Chan and Vese proposed a
simplified C-V model using the weighted mean of the gray scale as a constant to approximate
the image data values inside and outside the contour [30, 31]. Chen [5] proposed a region-
scalable fitting (RSF) model, which used two fitting energies (local gray mean values) to
approximate the grayscale of the image inside and outside the contour. At the same time, the
kernel function was introduced to improve the segmentation precision of low-contrast images
to some extent. Combined with other algorithm models, it had been well applied in medical
image segmentation, but overall, it was still sensitive to noise and contour initial position.
The level set method implicitly represents the evolution curve as a zero-level set
(hypersurface) of a high-dimensional level set function, and the evolution curve evolves under
the control of a partial differential equation (PDE) until its zero-level set corresponds to the
edge of the detected object. Compared with other deformation contour models, the level set
method is an implicit representation model that does not require curve or surface parameter-
ization, and the viscosity solution is obtained by the numerical scheme of partial differential
equations [23]. It has been used in computer vision and image processing, such as image
segmentation, image recovery, image enhancement, 3D image reconstruction.
Deep learning with convolutional neural networks (CNNs) has achieved state-of-the-art
performance for automated medical image segmentation. However, automatic segmentation

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

31052 Multimedia Tools and Applications (2020) 79:31049–31068

methods have not demonstrated sufficiently accurate and robust results for clinical use due to
the inherent challenges of medical images, such as poor image quality, different imaging and
segmentation protocols, and variations among patients. Gong [7] introduced top-k average
pooling to compute the channel co-attention descriptor. Then the correlation between features
in different spatial positions was explored to get the spatial co-attention descriptor. Finally,
these two types of co-attention descriptors were multiplied to generate a fused one. Such a
fused co-attention descriptor for each image was used to produce the co-attention augmented
feature map for the following processing in the applications. Li [2] proposed a degenerating
model algorithm that could simultaneously select the optimal number of mixture components
and estimate the parameters for Gaussian mixture model. Unlike traditional model order
selection method, it did not need to select the optimal number of components from a set of
candidate models. Based on the investigation on the property of the elliptically contoured
distributions of generalized multivariate analysis, it selected the correct model order in a
different way that needed less operation times and less sensitive to the initial value of EM.
Sun [21] proposed an enhanced affinity graph for image segmentation. And other researchers
proposed image segmentation methods [13, 17, 27].

3 Preliminaries

The famous C-V model is a simplification of the Mumford-Shah (M-S) model, eliminating the
complex mathematical description. Cluster analysis plays an important role in image process-
ing and is widely used in medical images and other fields. In order to make up for the
shortcomings of the two types of image processing tools, in the C-V model, combined with the
clustering idea, the segmentation process is regarded as a global density-oriented BIRCH
(balanced iterative reducing and clustering using hierarchies) clustering problem. That is, with
the given image pixel as data, the data is divided into two categories K = 2, which are the target
and background part to be searched respectively. For the gray-scale I(x) of a given image R, the
C-V model performs extreme optimization on the following formula.
2
ECV ðC; c1 ; c2 Þ ¼ ∑ λi ∫Ωi jI ðxÞ−ci j2 dx þ vjCj ð1Þ
i¼1

Where Ωi is the target and the background. ci is the gray mean value of the inner and outer
region of the contour C. The summation term is the external energy term of the active contour
model. The third term is the internal energy term, or penalty term. λ and ν are the factors. After
introducing the level set method, the contour C of the segmentation curve is replaced by the
zero level set of the level set function φ(x), then the C-V model for the level set function is
expressed as follows.
2
E CV ðφðxÞ; c1 ; c2 Þ ¼ ∑ λi ∫Ωi M i φðxÞjI ðxÞ−ci j2 dx þ v ∑ ðφðxÞÞj∇φðxÞjdx ð2Þ
i¼1 Ω

here Mi(⋅) is the Heaviside function, and φ(⋅) is the H function derivation (Dirac function) [18].
The third contour C and the membership function M are respectively expressed as gradients and
derivatives. The C-V model uses the global gray mean to segment the image, which greatly
improves the segmentation effect. However, for gray non-uniform images, especially medical
images, it has been interfered by noise, which seriously affects the segmentation accuracy.

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Multimedia Tools and Applications (2020) 79:31049–31068 31053

4 Proposed medical image segmentation model

The flowchart of proposed method is shown in Fig. 1, which contains four stages. Then we
detailed explain each process in this section.

4.1 Density-oriented BIRCH clustering and global energy

The BIRCH clustering [14] uses the pixel gray value of the image as the sample data, ignoring
the domain relationship between the pixels (the spatial information). According to the char-
acteristics of medical images, an adjacent pixel is likely to belong to the same class. For
medical images, the gray mean in a small neighborhood of each pixel is used, and a new image
is reconstructed to reduce the influence of noise on cluster segmentation. The neighborhood
size is usually set to 3 × 3, the size of the original image R(i, j) is M × N, and the new image
Q(i, j) is expressed as follows.

1 iþ1 jþ1
Qði; jÞ ¼ ∑ ∑ Rði; jÞ ð3Þ
M N k¼i−1 l¼ j−1

Where i = 1, 2, ⋯, M, j = 1, 2, ⋯, N. k and l form a neighborhood range. In this way, in the

process of segmenting the image, the combination of gray scale and spatial information can
effectively reduce the influence of noise.
Meanwhile, since the C-V model uses the mean square error as the extremum optimization
criterion, in order to further reduce the outlier interference, the K-means clustering algorithm
based on the correlation entropy is adopted, and the objective function is denoted as,
K  
min− ∑ ∑ uQ ð j; iÞσ2 g I Q ðxÞ−C i 2
ð4Þ
i¼1 x∈Ω

In here, Q is re-sampling image, g is the Gaussian kernel function to assist the K-means
clustering based on the correlation entropy to be closer to the cluster center. The selected
 
−kI Q ðxÞ−C i k2
2

Gaussian kernel function is: exp 2σ2 .

Selecting the continuous form of the correlation entropy as the global fitting term of the
energy functional, as follows:
K  
E CG ðc1 ; c2 ; ⋯; C K ; φÞ ¼ − ∑ λi ∫Ωi M i φðxÞg I Q ðxÞ−C i 2
dx ð5Þ
i¼1

Here, C1, ⋯, Ck are the respective cluster centers Ci. Through the adjustment and control of the
parameter k, the evolution curve searches for the target faster, and at the same time effectively
reduces the influence of the outliers.

DBIRCH Improved statistical

clustering active contours

Global energy Local energy

Input image ROI region Level Set Regularization Output result

Fig. 1 Flowchart of proposed segment method

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

31054 Multimedia Tools and Applications (2020) 79:31049–31068

4.2 Improved statistical active contours and local energy

Using the improved region-scalable fitting model, namely robust statistical information [25], a
new input image is formed to replace the original image R. Using the relevant knowledge of
the statistical distribution, the gray median MED, the interquartile range IQR and the average
absolute deviation MAD are used as the statistics of the image data. Let the gray value be I(xt)
of a pixel point, t is a certain pixel point in all pixel points n. Then the gray value vector is
(I(x1), I(x2), ⋯, I(xn)), xt ∈ (x1, x2, ⋯, xn), the statistics are denoted as follows:

xðnþ1Þ=2 ; n ¼ 2m þ
1
MEDðxÞ ¼ ð6Þ
0:5 xn=2 þ xðnþ1Þ=2 ; n ¼ 2m

IQRðxÞ ¼ Q3 ðxÞ−Q1 ðxÞ ð7Þ

jxt −xj
MADðxÞ ¼ ð8Þ
n
x is simplified to represent the corresponding gray value. Where Q is the upper/lower quantile
point and x is the gray mean value, then the new input image is expressed as:
I RS ðxÞ ¼ MEDðxÞ þ ρ1 IQRðxÞ þ ρ2 MADðxÞ ð9Þ

ρ is the information fusion coefficient, which is related to gray unevenness and noise. The
above formula embodies a good segmentation effect when dealing with grayscale inhomoge-
neities and noise images. But because it is a local method, it is sensitive to the initial position.
Therefore, combining it with the global energy of the previous section can effectively adjust
the ratio of global to local action. Local energy is written as follows:
 
EL f 1 ðxÞ; f 2 ðxÞ; ⋯; f K ðxÞ ; φ ¼
K ð10Þ
∑ λi ∫Ω M i φðxÞðI RS −f i ðxÞÞ2 dx
i¼1

Where (λ1, λ2, ⋯, λK) are positive constant coefficients corresponding to each cluster. (f1, f2,
⋯, fK) represents the gray mean of the local region based on the statistical information of the
original input image. The specific calculation is as follows:

∫σ ðx−yÞI RS ðyÞM 1 ðφðyÞÞdy

f 1 ð xÞ ¼ ð11Þ
∫σ ðx−yÞM 1 ðφðyÞÞdy

∫σ ðx−yÞI RS ðyÞM 2 ðφðyÞÞdy

f 2 ð xÞ ¼ ð12Þ
∫σ ðx−yÞM 2 ðφðyÞÞdy

∫σ ðx−yÞI RS ðyÞM K ðφðyÞÞdy

f K ð xÞ ¼ ð13Þ
∫σ ðx−yÞM K ðφðyÞÞdy

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Multimedia Tools and Applications (2020) 79:31049–31068 31055

For a two-phase level set, namely, only the target and background two clusters K = 2, the
equation can be simplified as:

E L ð f 1 ðxÞ; f 2 ðxÞ; φÞ ¼ ∫Ω M 1 ðφÞðI RS ðxÞ−T Þdx ð14Þ

f 1 ðxÞþ f 2 ðxÞ
Where T ¼ 2 , f ðiÞ ¼ K σ ½KMσi½ðMφiððxφÞÞIðxRSÞÞðxÞ ; i ¼ 1; 2.

4.3 Level set regularization

The penalty and regular terms in the energy formula of the RSF model [8] are re-presented as
follows:

E R ¼ μ∫Ω δðφðxÞÞj∇φðxÞjdx þ v∫Ω 0:5ðj∇φðxÞj−1Þ2 dx ð16Þ

At the same time, the smooth functions Hε and δε are simplified to H and δ, respectively.
 
2 φ
H ε ðφÞ ¼ 0:5 1 þ arctan ð17Þ
π ε

0 1 ε
δε ðφÞ ¼ H ε ðφÞ ¼ ð18Þ
π ε 2 þ φ2
So we get the final energy function of the proposed method.

E ¼ λE G þ ð1−λÞE L þ E R ð19Þ

Where λ ∈ [0, 1] is weighted positive coefficient to control the distribution ratio of global and
local energy, and its value is usually determined by gray unevenness. For grayscale uniform
images, it takes a small value, otherwise takes a larger value.

4.4 Energy minimization

The least squares or gradient descent method is usually used to solve the functional extremum.
In this paper, the gradient descent method is used. That is, the variational Euler equation is
transformed into a dynamic PDE problem [4]. When the evolution of the curve reaches steady
state, the solution of the variational Euler equation can be obtained. In numerical calculations,
discrete approximation calculations are usually performed using a finite difference method.
In this paper, the two-phase level set is taken as the research content, and the image data is
divided into two kinds of clusters, namely single target and background. The level set function
is kept as a simple binary step function instead of the symbol distance function.

5 Experiments and analysis

In order to verify the feasibility of the proposed algorithm, the improved model is tested on
medical images and compared to hybrid level set model (HLS) [6], hybrid level set model
based on LBF (HLS-LBF) [1] and three state-of-the-art segmentation methods: RAC [16], REI

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

31056 Multimedia Tools and Applications (2020) 79:31049–31068

[28] and HFC [34]. The selected experimental environment is Windows10 operating system,
GPU, Pentium(R) Dual-Core, 16G memory, MATLAB R2017a.
The parameters used in this paper are uniformly set to: σ = 5, ρ1 = ρ2 = 0.1, λ = 0.1, μ = 0.2,
v = 0.002 × 255 × 255, ε ¼ 1:5, k2 = 10, the template size of the Gaussian kernel function is
(3σ + 1) × (3σ + 1) according to extensive experiments.
In this paper, Dice similarity coefficient (Dice) and peak signal-to-noise ratio (PSNR) are
used for quantitative evaluation. Dice is denoted as follows:
2ðA∩BÞ
Dice ¼ ð20Þ
AþB
In this formula, A and B represent the gold standard and the area of the object area after
segment respectively. PSNR is expressed as follows:
 
d
PSNR ¼ 10log10 ð21Þ
MSEðA; BÞ

d is the highest grey level value in the image. MSE(A,B) is the mean square error of the two
regions.
We select four different retinal vessel images from DRIVE dataset as shown in
Fig. 2.
As can be seen from the figure, the four images all have obvious phenomenon of
uneven intensity, weak edge and low contrast. The improved model is used to segment
these images. The image size is 4000 × 3000pixel. Figure 3 is the enhanced gray image.
Figure 4 is the image after level set regularization. The segmented image is shown in Fig. 5
using proposed method.

(a) (b)

(c) (d)
Fig. 2 Four testing images in this study. a) and c) left retina vessel image; b) and d) right retina vessel image

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Multimedia Tools and Applications (2020) 79:31049–31068 31057

(a) (b)

(b) (d)
Fig. 3 Enhanced gray image

We mainly combine density-oriented BIRCH clustering (DBIRCHC) and active contour

method (ACM) to segment the medical images. Therefore, we firstly give the comparison
with/without the above two components as shown in Table 1.

(a) (b)

(b) (d)
Fig. 4 Image after level set regularization

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

31058 Multimedia Tools and Applications (2020) 79:31049–31068

(a) (b)

(c) (d)
Fig. 5 Segmented results with proposed method

Table 1 Comparison with/without DBIRCHC and ACM

Method With DBIRCHC and Without ACM Without DBIRCHC and With ACM Proposed

Index Dice PSNR Time Dice PSNR Time Dice PSNR Time

Image 1 63.2% 11.2 1.3 s 72.3% 16.5 0.8 s 97.5% 23.54 1.2 s
Image 2 65.1% 12.8 0.9 s 71.5% 18.7 0.7 s 97.4% 26.87 1.3 s
Image 3 61.8% 13.4 0.8 s 73.5% 19.2 0.8 s 96.5% 25.96 1.2 s
Image 4 62.7% 12.6 0.9 s 72.9% 16.7 0.6 s 97.3% 25.53 1.3 s

Because the proposed method combines the above two methods, the running time is
obviously slightly lower than them. However, the efficiency is superior to them.
Table 2 and Table 3 are the Dice and PSNR comparison results. Also the time comparison
is shown in Table 4. Obviously, The improved model obtains good segmentation results with
less false segmentation and more robust segmentation results.
We also use the true positive rate (TPR), false positive rate (FPR), and receiver
operating characteristic (ROC) as performance metrics. The result is shown in Fig. 6.

Table 2 Dice value with five methods

Method HLS/% HSL-LBF/% RAC/% HFC/% REI/% Proposed/%

Image 1 79.5 81.7 87.8 91.2 92.6 97.5

Image 2 80.8 86.5 89.2 90.8 93.5 97.4
Image 3 74.6 85.4 87.4 92.3 93.7 96.5
Image 4 79.3 82.6 88.5 91.7 94.1 97.3

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Multimedia Tools and Applications (2020) 79:31049–31068 31059

Table 3 PSNR value with five methods

Method HLS/% HSL-LBF/% RAC/% HFC/% REI/% Proposed/%

Image 1 18.92 19.23 19.65 21.38 22.64 23.54

Image 2 17.33 17.65 17.84 19.86 23.57 26.87
Image 3 18.21 18.46 18.65 20.67 23.87 25.96
Image 4 18.84 19.14 19.32 21.17 24.13 25.53

Table 4 Time comparison with five methods/s

Method HLS HSL-LBF RAC HFC REI Proposed

Image 1 3.1 2.7 2.3 2.1 1.8 1.2

Image 2 3.8 3.1 2.2 1.9 1.4 1.3
Image 3 2.9 2.4 2.2 2.1 1.6 1.2
Image 4 3.2 2.8 2.4 2.4 1.7 1.3

Fig. 6 ROC curves

Fig. 7 Flow chart of CNN-based methods

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

31060 Multimedia Tools and Applications (2020) 79:31049–31068

A larger interval between the TPR and FPR indicates a better segmentation performance.
In the comparison results, the performance of the HLS algorithm indicates that it is not
better for describing wispy features such as eye. Some algorithms, such as HSL-LBF and
RAC, do not express medical image characteristics well. HLS achieves a lower FPR, but
its lower TPR means that it is not sufficiently suitable for describing eye segment. The
algorithm proposed in this paper achieves a higher level of TPR than other methods.
Some schemes, such as HFC, are slightly better than our algorithm, but spend more time
in the feature extraction process.
Nowadays, deep learning methods are frequently used for image segment. To illus-
trate the robustness of proposed image segment method, we select other two deep
learning methods including FCN [10], JSLP [29] to make comparison. Common brain
diseases contain brain tumor, brain trauma, acute cerebrovascular diseases (cerebral
hemorrhage, cerebral infarction), brain atrophy, and their imaging manifestations are
different. Multiple MRI images of brain cases are selected from Harvard hologram,
including metastatic bronchial cancer, sarcoma, metastatic colon cancer, meningioma
and metastatic adenocarcinoma, T2-weighted images.

Fig. 8 Segmentation results for metastatic bronchial cancer. (a) original image, (b) Proposed method, (c) JSLP,
(d) FCN

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Multimedia Tools and Applications (2020) 79:31049–31068 31061

We used the Caffe library to implement FCN Net. The training process was done via one
node of the Emerald cluster3 with two 8-core E5-2623v3 Intel Haswells, a K80 NVIDIA GPU
and 128GB memory. To deal with different organs and different modalities, the region inside a
bounding box was normalized by the mean value and standard deviation of that region, and
then used as the input of the CNNs. In the training stage, the bounding box was automatically
generated based on the ground truth label with a random margin in the range of 0 to 10 pixels/
voxels. We used cross entropy loss function and stochastic gradient decent with momentum
0.9, batch size 1, weight decay 5 × 10−4, maximal number of iterations 80 k and initial learning
10−3 that was halved every 5 k iterations.
In the testing stage, the trained CNN models were deployed to a MacBook Pro (OS X
10.9.5) with 16GB RAM, an Intel Core i7 CPU running at 2.5GHz and an NVIDIA GeForce
GTX 1060 M GPU. A Matlab GUI and a PyQt GUI were used for user interactions on medical
images. In each network update step, we used a learning rate 10−2 and iteration number 20. We
used a grid search with the training data to get proper values of parameters, and fixed them as
global parameters during testing [26].
We summarize the flow chart of CNN-based methods as shown in Fig. 7:
During training, each instance is cropped with its bounding box, and the CNN is trained for
binary segmentation. In the testing stage, image-specific fine-tuning with optional scribbles

Fig. 9 Segmentation results for sarcoma. (a) original image, (b) Proposed method, (c) JSLP, (d) FCN

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

31062 Multimedia Tools and Applications (2020) 79:31049–31068

Fig. 10 Segmentation results for metastatic colon cancer. (a) original image, (b) Proposed method, (c) JSLP, (d)
FCN

and a weighted loss function is used. Note that the object class for testing may have not been
present in the training set.
The comparison results are shown in Figs. 8, 9, 10, 11 and 12.
In order to better describe the effectiveness of the proposed algorithm for process-
ing medical images, inter-class variance and image entropy are used for quantitative
analysis. Tables 5, 6, 7, 8, 9 show the segmentation effect with different algorithms.
The inter-class variance is used to judge the quality of segmented images according to
the contrast of inter-region characteristics. The inter-class variance can be calculated
from Eq. (21), where k is the number of clustering centers. N and M denote the
proportion of the first and second area, respectively (it is the number of pixels in the
adjacent areas). U1 and U2 are the average gray value of the first and second region.
U denotes the average gray value of the two regions. Inter-class variance and image
entropy can quantitatively describe the effect of image segmentation. If the value is
larger, the difference between the two categories is greater, so the image content is
more richer.
k−1 N M
SEC ¼ ∑ ðU 1 −U Þ2 þ ðU 2 −U Þ2 ð22Þ
i¼1 M þ N M þN

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Multimedia Tools and Applications (2020) 79:31049–31068 31063

Fig. 11 Segmentation results for meningioma. (a) original image, (b) Proposed method, (c) JSLP, (d) FCN

In this paper, proposed segmentation algorithm does not need prior knowledge when
aiming at the complex medical images with no obvious regional gray difference, the result
of segmentation is ideal. Where the segmented image by proposed algorithm has a higher
entropy than that of JSLP and FCN algorithm. The SEC is basically larger than that of the two
algorithms. This indicates that when the proposed algorithm segments gray images, it can
retain more complete information. The segmentation effect is better. In terms of computational
speed, due to no training process, the proposed algorithm requires few time than other two
methods.
We also conduct experiments on public medical dataset. 3D abdominal CT scans are
used in this experiment to evaluate the performance of the proposed method. Image
acquisition parameters and patient demographics for the dataset used here can be found
in [22]. Pre-processing and augmentation are carried out in similar manner to those for
sarcoma segmentation. The 128 × 128 image patches are randomly cropped to develop
the training set.
We use the same the experimental settings and CNN configurations as in the previous
experiments, so no parameters tuning is performed in this experiment. The purpose is to
validate the flexibility of the proposed method. Therefore, we only split the whole dataset into
a training set (100 subjects) and a separate testing set (100 subjects).JSLP, the FCN.

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

31064 Multimedia Tools and Applications (2020) 79:31049–31068

Fig. 12 Segmentation results for metastatic adenocarcinoma. (a) original image, (b) Proposed method, (c) JSLP,
(d) FCN

Table 5 Comparison on metastatic bronchial cancer

Method JSLP FCN Proposed

SEC 1321.5 1594.4 1786.3

Entropy 4.12 4.38 4.87
Time 48.6 s 32.1 s 1.6 s

According to the results (displayed in Table 10 and Fig. 13) they have achieved fairly good
segmentations in terms of kidneys, liver, and spleen. The FCN method is much faster than the
JSLP. The FCN is the state-of-the-art method based on deep CNNs. It is clear that the FCN

Table 6 Comparison on sarcoma

Method JSLP FCN Proposed

SEC 978.4 1762.5 2751.5

Entropy 4.22 4.47 4.96
Time 39.5 s 25.4 s 1.8 s

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Multimedia Tools and Applications (2020) 79:31049–31068 31065

Table 7 Comparison on metastatic colon cancer

Method JSLP FCN Proposed

SEC 1023.7 1471.6 1987.6

Entropy 3.59 4.53 4.86
Time 41.8 s 32.7 s 1.5 s

Table 8 Comparison on meningioma

Method JSLP FCN Proposed

SEC 1522.1 1687.3 2176.9

Entropy 4.15 4.38 4.92
Time 42.2 s 31.8 s 1.7 s

Table 9 Comparison on metastatic adenocarcinoma

Method JSLP FCN Proposed

SEC 874.6 1459.8 2265.7

Entropy 4.25 4.66 4.84
Time 42.3 s 33.1 s 1.5 s

achieves significantly better results in the pancreas segmentation. Furthermore the inference
time is significantly reduced. However, in terms of the other organs, namely the kidneys, liver,
and spleen, the FCN did not offer significant improvements. Our proposed method obtains the
better result.

6 Conclusions

This paper proposes a new medical image segmentation, which adopts density-oriented
BIRCH clustering method to modify active contour model and improve the robustness of
noise. Combined with the active contour algorithm and region horizontal set, the local entropy
and robust statistics information are fused into a new algorithm of global and local fitting
energy, which improves the robustness of noise and can well deal with the problem of gray
level inhomogeneity. Experimental results show that compared with the proposed method can
segment medical images more accurately and has better effect.

Table 10 Comparison with different methods

Method JSLP FCN Proposed

SEC 964.7 1958.4 2687.3

Entropy 4.53 4.82 4.97
Time 36.2 s 35.8 s 2.7 s

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

31066 Multimedia Tools and Applications (2020) 79:31049–31068

Fig. 13 The visual examples of abdominal multi-organ segmentations. The first column displays the original
images. The second column shows the manual references. The following columns demonstrate the segmentations
of the JSLP, the FCN, and the proposed method

Acknowledgements This research was funded by Heilongjiang Province science found for returnees (grant
number: LC2017027), Jiamusi University Science and Technology Innovation Team Construction Project (grant
number: CXTDPY-2016-3), Basic Research Project of Heilongjiang Province Department Of Education (grant
number:2016-kyywf-0547).

References

1. Bai X, Li C, Sun Q, et al. (2009) Contrast constrained local binary fitting for image segmentation,
International Symposium on Advances in Visual Computing. Springer-Verlag
2. Bo LI, Liu W, Dou L (2010) Unsupervised learning of Gaussian mixture model with application to image
segmentation. Chin J Electron 19(3):451–456
3. Bonini S, Caivano G, 2014 Development of a LGD model Basel2 compliant: a case study, Mathematical
and Statistical Methods for Actuarial Sciences and Finance
4. Boskovitz V, Guterman H (2002) An adaptive neurofuzzy system for automatic image segmentation and
edge detection. Fuzzy Syst IEEE Trans 10(2):247–262
5. Chen Y, Yue X, Xu RYD, Fujita H (2017) Region scalable active contour model with global constraint.
Knowl-Based Syst 120(C):57–73
6. Gan Y, Xia Z, Xiong J, Zhao Q, Hu Y, Zhang J (2015) Toward accurate tooth segmentation from computed
tomography images using a hybrid level set model. Med Phys 42(1):14–27
7. Gong X, Liu X, Li Y, Li H (2020) A novel co-attention computation block for deep learning based image
co-segmentation. Image Vis Comput 101:103973
8. Huo G, Yang SX, Li Q, Zhou Y (2016) A robust and fast method for Sidescan sonar image segmentation
using nonlocal Despeckling and active contour model. IEEE Trans Cybernetics 47(4):855–872
9. Koh J, Kim T, Chaudhary V, et al. 2010 Automatic segmentation of the spinal cord and the dural sac in
lumbar MR images using gradient vector flow field. Eng Med Biol Soc 3117–3120
10. Lee J, Nishikawa RM (2018) Automated mammographic breast density estimation using a fully
convolutional network. Med Phys 45(3):1178–1190

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Multimedia Tools and Applications (2020) 79:31049–31068 31067

11. Lin T, Li H, Yin S, Yang S (2019) Improved krill group-based region growing algorithm for image
segmentation. Int J Image Data Fusion 10:327–341. https://doi.org/10.1080/19479832.2019.1604574
12. Liu J, Yin S-L, Li H, Lin T (2017) A density-based clustering method for K-anonymity privacy protection. J
Inf Hiding Multimed Signal Process 8(1):12–18
13. Lu X, Wang W, Ma C, et al. (2019) See more, know more: unsupervised video object segmentation with co-
attention Siamese networks [C]// CVPR19
14. Madan S, Dana KJ (2016) Modified balanced iterative reducing and clustering using hierarchies (m-
BIRCH) for visual clustering. Pattern Anal Appl 19(4):1023–1040
15. Miao J, Gu X, Ying H et al (2018) Medical Imaging. Int Topical Meeting Image Detect Qual 80(5):1–4
16. Niu S, Chen Q, Sisternes LD et al (2017) Robust noise region-based active contour model via local
similarity factor for image segmentation. Pattern Recogn 61:104–119
17. Perazzi F, Khoreva A, Benenson R, et al. (2017) Learning video object segmentation from static images
[C]// 2017 IEEE conference on computer vision and pattern recognition (CVPR). IEEE Computer Society
18. Petke J, Jeavons P (2010) Local consistency and SATSolvers. J Artif Intell Res 43(3):398–413
19. Qi G, LongW SS (2015) Multiple-Channel local binary fitting model for medical image segmentation. Chin
J Electron 24(4):802–806
20. Singh E, Lin D, Barrett C, et al. 2018 Logic bug detection and localization using symbolic quick error
detection, IEEE transactions on computer-aided Design of Integrated Circuits and Systems, PP(99):1–1.
21. Sun G, Lin K, Wang J et al (2019) An enhanced affinity graph for image segmentation. ICE Trans Inf Syst
E102–D(5):1073–1080
22. Tong T, Wolz R, Wang Z, Gao Q, Misawa K, Fujiwara M, Mori K, Hajnal JV, Rueckert D (2015)
Discriminative dictionary learning for abdominal multiorgan segmentation. Med Image Anal 23(1):92–104
23. 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
24. Wang HJ, Liu M, Ma WL 2010 Color image segmentation based on a new geometric active contour model.
Int Conf Mach Vision Human-Mach Interface
25. Wang J, Ju L, Wang X (2016) An edge-weighted Centroidal Voronoi tessellation model for image
segmentation. Comput Mathematics Appl 71(11):2272–2284
26. Wang G, Li W, Zuluaga MA, Pratt R, Patel PA, Aertsen M, Doel T, David AL, Deprest J, Ourselin S,
Vercauteren T (2018) Interactive medical image segmentation using deep learning with image-specific fine
tuning. IEEE Trans Med Imaging 37(7):1562–1573
27. Wang W, Lu X, Shen J, Crandall D, Shao L, Zero-Shot Video Object Segmentation via Attentive Graph
Neural Networks 2019 IEEE/CVF international conference on computer vision (ICCV), Seoul, Korea
(South): pp 9235–9244, doi: https://doi.org/10.1109/ICCV.2019.00933.
28. Wang X, Li W, Zhang C, Lou W, Song R (2019) An adaptable active contour model for medical image
segmentation based on region and edge information. Multimed Tools Appl 78:33921–33937. https://doi.
org/10.1007/s11042-019-08073-3
29. Yan Z, Yang X, Cheng K-T (2018) Joint segment-level and pixel-wise losses for deep learning based retinal
vessel segmentation. IEEE Trans Biomed Eng 65(9):1912–1923
30. Yin S, Zhang Y (2018) Singular value decomposition based anisotropic diffusion for fusion of infrared and
visible images. Int J Image Data Fusion 9(4):1–18
31. Yin SL, Zhang Y, Karim S (2018) Large scale remote sensing image segmentation based on fuzzy region
competition and Gaussian mixture model. IEEE Access 6:26069–26080
32. Zhang L, Jiang M, Yang M 2010 Hippocampus segmentation based on prior knowledge and GAC model.
IEEE Int Conf Signal Process
33. Zhang K, Zhang L, Lam KM, Zhang D (2016) A level set approach to image segmentation with intensity
inhomogeneity. IEEE Trans Cybernetics 46(2):546–557
34. Zhang M, Jiang W, Zhou X et al (2017) A hybrid biogeography-based optimization and fuzzy C-means
algorithm for image segmentation. Soft Comput 1:1–14

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

31068 Multimedia Tools and Applications (2020) 79:31049–31068

Affiliations

Shoulin Yin 1 & Hang Li 1 & Desheng Liu 2 & Shahid Karim 3
1
Software College Shenyang Normal University, Shenyang 110034, China
2
College of Information and Electronic Technology, Jiamusi University, Jiamusi 154007 Heilongjiang,
China
3
Department of Computer Science, ILMA University, Karachi, Pakistan

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center
GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers
and authorised users (“Users”), for small-scale personal, non-commercial use provided that all
copyright, trade and service marks and other proprietary notices are maintained. By accessing,
sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of
use (“Terms”). For these purposes, Springer Nature considers academic use (by researchers and
students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and
conditions, a relevant site licence or a personal subscription. These Terms will prevail over any
conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription (to
the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of
the Creative Commons license used will apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may
also use these personal data internally within ResearchGate and Springer Nature and as agreed share
it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not otherwise
disclose your personal data outside the ResearchGate or the Springer Nature group of companies
unless we have your permission as detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial
use, it is important to note that Users may not:

1. use such content for the purpose of providing other users with access on a regular or large scale
basis or as a means to circumvent access control;
2. use such content where to do so would be considered a criminal or statutory offence in any
jurisdiction, or gives rise to civil liability, or is otherwise unlawful;
3. falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association
unless explicitly agreed to by Springer Nature in writing;
4. use bots or other automated methods to access the content or redirect messages
5. override any security feature or exclusionary protocol; or
6. share the content in order to create substitute for Springer Nature products or services or a
systematic database of Springer Nature journal content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a
product or service that creates revenue, royalties, rent or income from our content or its inclusion as
part of a paid for service or for other commercial gain. Springer Nature journal content cannot be
used for inter-library loans and librarians may not upload Springer Nature journal content on a large
scale into their, or any other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not
obligated to publish any information or content on this website and may remove it or features or
functionality at our sole discretion, at any time with or without notice. Springer Nature may revoke
this licence to you at any time and remove access to any copies of the Springer Nature journal content
which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or
guarantees to Users, either express or implied with respect to the Springer nature journal content and
all parties disclaim and waive any implied warranties or warranties imposed by law, including
merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published
by Springer Nature that may be licensed from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a
regular basis or in any other manner not expressly permitted by these Terms, please contact Springer
Nature at

[email protected]