Microwave Medical Image Segmentation for Brain

Stroke Diagnosis: Imaging-Process-Informed Image

2023 IEEE 17th International Symposium on Medical Information and Communication Technology (ISMICT) | 979-8-3503-0417-6/23/$31.00 ©2023 IEEE | DOI: 10.1109/ISMICT58261.2023.10152073

Processing
Chenghui Liu Zheng Gong*
School of Life Science and Technology School of Engineering
University of Electronic Science and Technology of China The University of Waikato
Chengdu, China Hamilton, New Zealand
[email protected] [email protected]

Yifan Chen* Shuaiting Yao

School of Life Science and Technology School of Life Science and Technology
University of Electronic Science and Technology of China University of Electronic Science and Technology of China
Chengdu, China Chengdu, China
[email protected] [email protected]

Abstract—In this paper, we propose a novel imaging-process- imaging (MMI) uses microwave radiation to extract internal tis
informed image segmentation method that accounts for sue information and reconstruct the spatial distribution of diele
uncertainty during the imaging process. A priori information is ctric properties using a discrete element model [6], [7]. MMI ha
incorporated to enhance the contrast between stroke area and s emerged as a potential alternative method for detecting stroke
healthy tissues. The distorted Born iterative method (DBIM) is , and has a wide range of applications in civil, industrial, and m
utilized to reconstruct the stroke area of the brain. Due to the non- edical fields, including non-destructive testing, materials charac
linear relationship between actual and estimated dielectric terization, remote sensing, and medical diagnostics [8]. Additio
constants resulting from DBIM, the microwave medical image nally, the use of low-power, non-ionizing microwave radiation
lacks a clearly defined boundary, posing a challenge to accurately
in MMI poses minimal risk to human health, and MMI systems
segment it using traditional methods. The proposed method
achieves effective image segmentation by improving the
can be designed to be portable and inexpensive [8]. The distort
traditional threshold method. From the simulation results, the ed Born iterative method (DBIM) is a widely used quantitative
region misclassified by the traditional method accounts for 89%, reconstruction algorithm that has been developed for solving th
while the proposed method results in a misclassification rate of e backscatter problem in microwave imaging [9].
only 13%. The results demonstrate a significant improvement of Image segmentation plays a crucial role in extracting useful
58.85% in accurately reproducing the dielectric constants. information from the data obtained through various medical im
aging modalities. Therefore, understanding the different image
Keywords—Imaging-process-informed imaging processing,
image segmentation, microwave medical imaging, distorted Born
segmentation techniques is essential for achieving accurate and
iterative method, brain stroke diagnosis. reliable results in MMI and other imaging modalities. Image se
gmentation methods can be broadly categorized into several cat
I. INTRODUCTION egories, such as thresholding and region-growing approaches. T
hresholding method segments scalar images by creating a binar
Stroke remains a major global health challenge, ranking sec y partition of image intensities, where a threshold is determined
ond among the leading causes of death worldwide, and first am to separate desired classes. Region-growing method is a techni
ong the leading causes of disability [1]. Traditional imaging tec que for extracting a connected image region based on predefine
hniques for stroke detection, such as computed tomography (C d criteria [10]. However, the images produced by MMI using D
T) and magnetic resonance imaging (MRI), play a critical role i BIM often lack clarity due to the inherent uncertainty during th
n diagnosis and treatment [2], [3]. However, their use in comm e imaging process. Unlike other imaging methods (e.g., MRI) t
unity hospitals and general clinics is limited due to their large s hat can obtain clear boundaries, microwave imaging can only p
ize and high cost [4]. Furthermore, the lack of portability of CT roduce a “continuous” image, which results in the lack of sharp
and MRI devices and their long imaging time pose a risk of ex ness for medical images [11], [12]. Specifically, the non-linear
acerbating symptoms in stroke patients [5]. Microwave medical relationship during the imaging process using DBIM, i.e., the s

Authorized licensed use limited to: University of Sargodha. Downloaded on November 04,2023 at 16:27:40 UTC from IEEE Xplore. Restrictions apply.
olution of the Green’s function, hinders the accurate acquisition
of dielectric constant values [13], [14]. Due to the indefinite bo
undaries of MMI images, traditional medical image segmentati
on methods mentioned above cannot obtain effective segmentat
ion results. Some research works on MMI image segmentation
divide images into different tissue types, allowing for a quantita
tive evaluation of regions containing specific tissues [14], [15].
However, these approaches do not take into account the uncerta
inties introduced by the imaging process (i.e., the non-linear rel
ationship caused by MMI) to improve the performance of imag
e segmentation.
To overcome the above problem, this paper advances the
state-of-the-art by developing an imaging-process-informed
image segmentation (IPI-IS) method that considers uncertainty
during the imaging process (i.e., DBIM). Besides, a priori
knowledge based on physiological models is also incorporated
to enhance the contrast between hemorrhagic regions and
healthy tissues. With the proposed method, we first estimate the
non-linear relationship caused by DBIM using statistical
learning methods, and then use this information to adjust the
image for a more accurate determination of the dielectric
constants. Finally, the accurate image segmentation of brain
stroke is achieved. Fig. 1. Algorithm flowchart.

The remaining part of this paper is organized as follows: based on the above sensing curve. Due to the statistical nature
Section II presents a general description of IPI-IS algorithm. of the sensing curve, it may not cover the entire range of
Section III applies the proposed algorithm to a numerical brain dielectric constants for the given layer, leading to handling
phantom for system verification. Finally, Section IV difficulties for special values. Therefore, the adjustment curve
summarizes this paper and gives some future works. can be obtained by introducing API and sensing curve.
Specifically, we utilize API to determine the range of dielectric
II. IMAGING-PROCESS-INFORMED IMAGE SEGMENTATION constants for the given layer. Based on this range and the sensing
METHOD curve, we define the adjustment ratio for the layer. For parts
Conventional medical image segmentation methods are outside the range covered by the sensing curve, we use
ineffective in producing accurate segmentation results for MMI extrapolation to extend the adjustment curve. Finally, this layer
images with blurry boundaries. In this paper, we propose a is reconstructed based on the two curves to enhance the contrast
method for improving image segmentation by incorporating between the hemorrhagic area and healthy tissue. The Otsu
uncertainty in the imaging process. To account for the threshold obtained from the initial segmentation is then
uncertainty introduced by the DBIM imaging process, we use a reclassified based on the adjusted image. Through the above
statistical method to obtain the sensing curve. This curve cycle, we ultimately achieve multi-layer segmentation of the
represents the relationship between the actual and estimated image. The flow chart in Fig. 1 illustrates the proposed image
dielectric constants. To utilize this uncertainty more effectively, segmentation algorithm. In the following part, we will describe
we incorporate a prior information (API), such as the dielectric the algorithm using the first layer of RUE as an example.
constant range of the brain, to derive an adjustment curve. The The purpose of the proposed partitioning algorithm is to
adjustment curve is then used to modify the dielectric constant distinguish the first layer of the RUE from other layers as much
as possible. Let = {1, 2, … , } denote a digital image ( , )
of the image. The Otsu algorithm, a widely used tool for image
segmentation, is improved by incorporating the nonlinearity of
DBIM [10]. obtained by DBIM, with size × pixels. is the dielectric
constant range of RUE, where is the maximum dielectric
For microwave medical images, the region under constant in RUE. ( , ) is the gray value image, and each
examination (RUE) is typically a complex structure, such as the position in RUE has corresponding gray value, which
brain, which requires a layer-by-layer segmentation for represents dielectric constant. The nonlinear relationship
improved effective and accuracy. Assuming that there are induced by DBIM is estimated using statistical method to
layers in the RUE, the segmentation process requires − 1 obtain the sensing curve. Through simulation, we obtain a set
cycles. Each cycle consists of two curves for the dynamic of actual and estimated dielectric constant values corresponding
process to improve the performance of image segmentation: to the same brain location. Define the actual and estimated
, respectively, where =
sensing and adjustment curves. For each layer of the cycle, we
dielectric constant as and
incorporate DBIM imaging uncertainty to obtain the sensing
curve, which represents the corresponding relationship between [ , , . . , ] , = [ , , . . , ]. In this paper, we utilize a
the actual and estimated dielectric constants. This sensing curve linear relationship to describe the correspondence between the
is obtained by statistical method. To achieve more accurate actual and estimated dielectric constants:
determination of the dielectric constant, the image is adjusted

Authorized licensed use limited to: University of Sargodha. Downloaded on November 04,2023 at 16:27:40 UTC from IEEE Xplore. Restrictions apply.
 Fig. 3. Diagrammatic sketch of simulation; (b) is the image after
Fig. 2. Diagrammatic sketch of simulation. (a) and (b) are two (a) imaging, units in mm. Color bar: dielectric constant.
brain models with stroke area, units in mm. Color bar: dielectric
constant.
And
*(
0 ∪0 = , ). (6)
= ( , ), (1)
To evaluate the performance, a rubric is introduced.
where is the relationship between the actual and estimated Thresholding can be viewed as a statistical decision theory
dielectric constants caused by DBIM, is the regression problem that aims to maximize the difference between different
coefficient vector, = [! , ! , . . , ! ]. Then, the least squares groups when assigning pixels to two groups. Here, 56
method is used to implement curve fitting: represents the average dielectric constant of the whole image,
5$ is the average dielectric constant of class 0$ , 7$ is the
" = ∑$% [ $ − ( $ , )] . (2) probability that a pixel is assigned to 0$ . The deviation can be
However, it is obviously impractical to use a linear equation to expressed as:
89 = ∑$% 7$ (5$ − 56 ) .
approximate the non-linear relationship of DBIM. divide The
(7)
RUE is divided into multiple layers, each with a corresponding
range of dielectric constants. For each range, a linear By varying the threshold -, the maximum deviation can be
relationship is fitted to reconstruct the non-linear relationship obtained, leading to the optimal threshold - * :
of DBIM. To restore the real dielectric constants as in (1), we
introduce the magnitude of the dielectric constants of RUE by 89 (- * ) = max 89 (-), (8)
=>?>@
API. For different positions of the RUE, the dielectric constants
are resized and these coefficients constitute the adjustment and the first layer of the image can be segmented as follows:
curve. Assume that $ represents the magnitude of the dielectric 1, ′( , ) > - *
constant at a position in ( , ), and & is the adjustment ratio of ′( , ) = B . (9)
0, ′( , ) ≤ - *
this position:
By following the above process, we achieved the optimal
'(
&= . (3) segmentation of the first layer. The same approach can be
)('( , )
applied to segment the remaining layers until the entire image is
between this layer and other areas of the RUE. As a result, the segmented.
output image can be expressed as
III. RESULTS AND DISCUSSION
*
( , ) = ( , ) • &( , ). (4) In the above section, we have proposed a general IPI-IS
The above process describes the way to enhance the contrast method. In this section, the brain stroke is discussed as a use case
between the first layer and the other areas of RUE. Next, the to validate the performance of the proposed lgorithm. For
image segmentation is performed for the first layer by using the cerebral hemorrhage, the brain is the RUE, and it is generally
Otsu algorithm with the improved image. Suppose a dielectric divided into four layers, including skull, white matter, gray
constant threshold value is chosen as ,(-) = -, 0 < - < . matter, and stroke area. Therefore, N = 4. The Zubal Phantom is
Using this value to divide the input image into two categories, utilized as the simulation model [16] [17]. As shown in Fig. 2,
the first layer 0 and the other areas in RUE 0 . 0 comprises all brain model (a) is utilized as the training set to obtain the sensing
the pixels in the image with dielectric constant values within the and adjustment curves, whereas brain model (b) is utilized as the
range [1, -], while 0 consists of all the pixels with dielectric testing set. Using the statistical method proposed in Section II,
constant values within the range [- + 1, ] . Both 0 and 0 the sensing function can be obtained by changing the locations,
should satisfy the following conditions: the partition is sizes and dielectric constants of the stroke area using brain
complete, each pixel is assigned to a unique region, and the model (a). The imaging result of DBIM, as shown in Fig. 3, is
regions are disjoint: used to validate the algorithm, and the training set contains 30
images. Due to the nonlinearity of DBIM, one actual dielectric
0 ⋂ 0 = ∅, (5) constant results in multiple estimated values, and the average
value of all post-sensing dielectric constants is chosen as the

Authorized licensed use limited to: University of Sargodha. Downloaded on November 04,2023 at 16:27:40 UTC from IEEE Xplore. Restrictions apply.
 TABLE I. API OF TISSUE DIELECTRIC RANGES
Different structures of the
Dielectric constant
brain
Skull 3-20
White matter 30-45
Gray Matter 45-60
Stroke area 60-70 Fig. 6. (a) The original image, and (b) the adjusted image.

60

55

50

45

40
Measured Data
Curve-Fitting
35
45 50 55 60
Pre-sensing dielectric constants
Fig. 4. Sensing curve.

Fig. 5. Adjustment curve.

final post-sensing dielectric constant. Using statistical methods,

the different dielectric constant fixed intervals in (a) are
measured separately, and the range of the pre-imaging dielectric
constants is [45,60], and we obtain the post-imaging dielectric
constants for 10 groups. According to (1) and (2), the sensing
curve is statistically obtained. Due to the space limitation, we
have selectively displayed a portion of the results. Specifically,
Fig. 4 showcases the outcomes of the third layer. Based on API,
we know the dielectric constant ranges of the four layers in
advance, as shown in Table I. Based on the API and sensing
curve, the adjustment curve of the third layer is obtained. In
particular, as shown in Fig. 5, the adjustment curve extends from
interval [45, 60] to [35, 65]. Fig. 6 illustrates the adjusted image
obtained by applying the proposed algorithm. It is evident that
the adjusted image presents more prominent features and a more Fig. 7. The image segmentation by (a) the traditional Otsu
accurate restoration of dielectric constants, which facilitates the method and (b) the proposed method.
improvement of image segmentation accuracy. Fig.7 shows the
performance comparison between the traditional Otsu method red circle. The results indicate that the proposed method not only
and the proposed method. The stroke area is highlighted by the accurately restores the dielectric constants, but also performs

Authorized licensed use limited to: University of Sargodha. Downloaded on November 04,2023 at 16:27:40 UTC from IEEE Xplore. Restrictions apply.
more precise image segmentation of the stroke area. In order to [2] M. Giroud, A. Jacquin, Y. Béjot, “The worldwide landscape of stroke in
quantify the performance improvement of the proposed method, the 21st century,” Lancet, vol. 383, no. 9913, pp. 195, 2014.
we utilize two evaluation criteria: the segmentation [3] D. J. Mollura, M. P. Lungren, “Radiology in global health: strategies,
implementation, and applications,” Springer, 2014.
performance, and the accuracy of the reduction of the dielectric
[4] L. Baker, S. Atlas, C. Afendulis, “Expanded use of imaging technology
constants. The performance of the segmentation is reflected by and the challenge of measuring value,” Health Aff, vol. 27, no. 6, pp.1467-
calculating the ratio of the area of the segmented stroke region 1478, 2008.
to the real region. The misclassification rate of the traditional [5] J. L. Saver, “Time is brain: quantified,” Stroke, vol. 37, no. 1, pp. 263-
method accounts for 89%., while the proposed method results in 266, 2006.
a misclassification rate of only 13%. The reduction accuracy of [6] Z. Gong, Y. Chen, X. Lin, M. J. Cree, “Contrast-enhanced microwave
the actual and estimated dielectric constant refers to the cancer detection using angle-of-arrival approach,” IEEE Trans. Antennas
deviation: Propag., vol. 70, no. 5, pp. 3772-3780, 2021.
[7] C. Qin, Z. Zhao, Y. Li, “Recent advances in microwave imaging for breast
|G0 HIJ KG0 LM |
E= , (11) cancer detection,” IEEE Access, vol. 6, pp. 55885-55906, 2018.
G0 LM
[8] B. J. Mohammed, A. M. Abbosh, S. Mustafa, “Microwave system for he
ad imaging,” IEEE Trans. Instrum. Meas., vol. 63, no. 1, pp. 117-123, 20
where G0NOP and G0NQRS are the actual and estimated dielectric 13.
constants. The traditional method has a reduction of 5.93% [9] F. Gao, B. D. Van Veen, S. C. Hagness, “Sensitivity of the distorted born
while the new method is 2.44%. The performance improvement iterative method to the initial guess in microwave breast imaging,” IEEE
Trans. Antennas Propag., vol. 63, no. 8, pp. 3540-3547, 2015.
over the traditional method is 58.85%.
[10] D. L. Pham, C. Xu, J. L. Prince, “Annual review of biomedical
engineering”, Springer, 2000.
IV. CONCLUSION
[11] J. A. Martinez-Lorenzo, J. L. Gomez-Tornero, C. R. Salinas, “Microwave
In this paper, we propose a new imaging-process-informed imaging for medical applications: Current techniques, challenges, and
image segmentation method. Uncertainty in the imaging process future trends,” IEEE Trans. Microw. Theory Techn., vol. 65, no. 7, pp.
is introduced into the image segmentation. In addition, a priori 2387-2403, 2017
information is incorporated to enhance the contrast between the [12] C. Qin, Y. Li, “Review of microwave imaging based on the Born and Ry
tov approximations: When to use which?,” J. Electromagn. Waves App.,
stroke are and the healthy tissue. With the proposed method, we vol. 31, no. 5, pp. 463-481, 2017.
estimate the nonlinear relationships induced by DBIM using [13] P. M. Van den Berg, H. R. Kroeze, P. W. Verbeek, “Comparison of linear
statistical method, and then use this information to adjust the and non-linear image reconstruction algorithms for microwave breast
images to determine the dielectric constant more accurately. imaging,” Med. Biol. Engin. Com., vol. 53, no. 5, pp. 391-402, 2015.
Finally, the effective and accurate image segmentation of stroke [14] B. De Smedt, L. Crocco, D. J. Hawkes, “Comparison of different inversio
area is achieved. The experimental result shows that the n strategies for microwave breast imaging,” IEEE Trans. Med. Imag., vol.
reduction of the dielectric constant and the segmentation effect 36, no. 9, pp. 1835-1846, 2017.
are greatly improved compared with the traditional methods. For [15] X. Zhang, P. M. Meaney, “A method for tissue segmentation and bounda
ry detection in microwave imaging,” IEEE Trans. Biomed. Engin., vol. 5
future work, we will quantify and discuss the nonlinearity of 6, no. 10, pp. 2429-2436, 2009.
DBIM in more detail to further optimize the algorithm.
[16] I. G. Zubal, C. R. Harrell, E. O. Smith, Z. Rattner, G. Gindi, and P. B.
Hoffer, “Computerized three-dimensional segmented human anatomy,”
REFERENCES Med. Phys., vol. 21, no. 2, pp. 299-302, 1994.
[17] Z. Gong, Y. Chen, X. Lin, K. Wu, M. J. Cree, “Generic wideband
[1] V. L. Feigin, B. Norrving, G. A. Mensah, “Global burden of stroke,” phantom design methodology for microwave medical applications,” IEEE
Circul. Res., vol. 120, no. 3, pp. 439-448, 2017. Antennas Wireless Propag. Lett., vol. 20, no. 8, pp. 1488-1492, 2021.

Authorized licensed use limited to: University of Sargodha. Downloaded on November 04,2023 at 16:27:40 UTC from IEEE Xplore. Restrictions apply.