You are on page 1of 7

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/332029758

Improvement of MRI images through heterogeneous interpolation techniques

Conference Paper · March 2019


DOI: 10.1109/CONIELECOMP.2019.8673230

CITATIONS READS
0 61

4 authors, including:

Dante Mújica-Vargas
Centro Nacional de Investigación y Desarrollo Tecnológico
23 PUBLICATIONS   95 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Dante Mújica-Vargas on 27 March 2019.

The user has requested enhancement of the downloaded file.


Improvement of MRI images through heterogeneous
interpolation techniques
1st Dante Mujica-Vargas
´ 2nd Yuridia Mújica-Vargas 3rd Manuel Matuz Cruz
Depto. de Ciencias Computacionales Subdirección de Prevención Depto. de Ciencias Computacionales
CENIDET-TecNM y Protección a la Salud CENIDET-TecNM
Cuernavaca, México Dirección Médica-ISSSTE Cuernavaca, México
dantemv@cenidet.edu.mx Ciudad de México, México mjmatuz@cenidet.edu.mx
ymujica@issste.gob.mx

3rd Arturo Rendón-Castro


Depto. de Ciencias Computacionales
CENIDET-TecNM
Cuernavaca, México
angel.rendon18ce@cenidet.edu.mx

Abstract—This paper presents the heterogeneous implementa- the process of estimating the intermediate values of a function
tion and evaluation of five algorithms to improve the resolution in or signal sampled in continuous positions or the attempt to
magnetic resonance images, these are Nearest-Neighbor, Bilinear, reconstruct the original continuous function from a set of
Bicubic, Catmull-Rom and Lanczos. The experiments were carried
out using the database BraTS 2017 considering T 1, T 2, T 1ce and discrete samples [10]. In the context of 2D signals (images),
F lair acquisition modes. To measure the efficiency and quality interpolation refers to the technique of recovering a continuous
of the results, the metrics PSNR, MAE, SSIM, as well as the signal by estimating the image data of a set of discrete samples
Runtime were used. The results show that for all modalities of image data. This task is fundamental in the processing of
the Bicubic algorithm has a P SN R > 35, M AE < 1.089, images, for which it is widely studied and applied, e.g. in [11],
SSIM > 0.958 and RT < 5.205 seconds per whole database.
By other hand, Catmull-Rom achieves favorable results compared it was proposed a video resolution improvement technique for
with Bicubic algorithm, maintaining the balance between quality low resolution video sequences based on a new directional
and a less processing time than mentioned method. interpolation of edge and complex wavelet transforms. In [12],
Index Terms—Brain MRI images, Interpolation techniques, the interpolation was used to reconstruct the satellite video
Heterogeneous system images. In [13], [14], were proposed new super-resolution
methods based on interpolation, the first work was based on
I. I NTRODUCTION an interpolation that considers the intensity distances and the
Computed tomography (CT), magnetic resonance imaging geometry of the pixel data, it was able to produce good texture
(MRI), and positron emission tomography (PET) are sophis- details; by other hand, the second one was based on fractional
ticated modalities typically used in the second-line diagnosis order gradient interpolation and let preserve the edges of the
following routine clinical practice [1]. Both in the diagnosis image.
processes and treatment planning, as well as in the control and In the context of medical images, the following researches
prediction of diseases [2]–[4] such as: cancer, cardiovascular can be mentioned: in [15], it was proposed a geometrical
diseases, trauma, neurological conditions as studied in [5], [6]. approach of the linear interpolation Nearest-Neighbor in order
In this regard, medical imaging has improved the accuracy, to accelerate the reconstruction of the HARDI signal from
timeliness and effectiveness of diagnoses. The majority of a few images of diffusion MRI (Diffusion Weighted MRI),
investigations in medical images refer to MR images due to the results were outstanding in terms of the quality of the
their flexible and non-invasive characteristics. These features data and a significant reduction in scan time. In [16], it was
make MRI an important diagnostic imaging technique, being proposed an approach based on restricted inverse diffusion
a valuable tool in the early detection of abnormal changes for the interpolation of 3D brain MR images, the proposal
in tissues and organs. However, this type of images regularly was compared with the linear interpolation methods, based
present a limited spatial resolution [7], [8]. on registration and CubicB-Spline, obtaining better results in
The spatial resolution of the image is an index to measure volumes of coarse slices. The approach proposed by [17] was
the clarity of the details of the image; improving the spatial based on an interpolation framework that considers 2 features
resolution of the image implies reducing the spectral alias (eigenvalues and orientation) in MRI images with diffusion
of the image by increasing the sampling frequency without tensor (Diffusion tensor MRI) of the human heart, the results
damaging the signal-to-noise ratio [9]. One way to perform this of this paper offered flexible options for interpolation while
procedure is using interpolation techniques. Interpolation is preserving the properties of diffusion tensors. Almost all image

978-1-7281-1145-2/19/$31.00 ©2019 IEEE 112


processing programs implement some interpolation technique C. Bicubic
for transformations, rotations and many other manipulations The 2D kernel required to develop the Bicubic interpolation
performed on an image. It is very important that the inter- is given by:
polation techniques have a low computational cost, which is
reflected in low processing time and high performance, that Wbic (x, y) = Wcub (x) · Wcub (y), (4)
is, more accurate results given by a significant improvement
in the fine details of the images. For example, in the field of where Wcub (x) and Wcub (y) can be computed as follow:
medical images, where the regions of the different tissues, as

3 2
|x| − 2 · |x| + 1
 for 0 ≤ |x| < 1
well as the abnormalities, are expected to be the best delimited, 3 2
Wcub (x) = −|x| + 5 · |x| − 8 · |x| + 4 for 1 ≤ |x| < 2 ,
in order to be able to segment, classify or perform a task of 
0 for |x| ≥ 2

higher hierarchy. In this regard, it is in our interests make
an analysis of classic interpolation techniques, implemented (5)
in a heterogeneous system (CPU + GPU), which allows to

3 2
|y| − 2 · |y| + 1
 for 0 ≤ |y| < 1
accelerate the process to be used in the treatment of complete
Wcub (y) = −|y|3 + 5 · |y|2 − 8 · |y| + 4 for 1 ≤ |y| < 2 ,
studies or real databases. 
0 for |y| ≥ 2

The rest of this paper is organized as follow. In Section II, a
briefly discussion about the evaluated interpolation algorithms (6)
is presented. The interpolation via separable convolutions is The resulting kernel is illustrated in the Fig.1(c). Due
introduced in Section III. The heterogeneous implementation to the decomposition in 1D kernels, the calculation of the
of the evaluated algorithms is stated in Section IV. Experimen- interpolation Bicubic is separable in both x and y [18].
tal results and a comparative analysis are presented in Section D. Catmull-Rom
V. In the final section the Conclusions are drawn and future
work is outlined. The cubic interpolation kernel described in previous section
(equations 5 and 6) are cubic polynomial functions by parts,
II. E VALUATED ALGORITHMS also known in the area of graphic computing as cubic splines.
In their general form, these not only take into account a single
A. Nearest-Neighbor
parameter a but also incorporate a second parameter b. If the
The position (ux , vy ) of the closest pixels to the interest parameters are fixed, e.g. a = 0.5 y b = 0, it is possible to
pixel (x, y) is determined by independently rounding the define a new interpolation technique so-called Catmull-Rom
coordinates x and y to discrete values, i.e. : spline, which one mathematically is expressed as
˜ y) = I(ux , vy ),
I(x, (1)

3 · |x|3 − 5 · |x|2 + 2 for 0 ≤ |x| < 1
1
where ux = round(x) = bx + 0.5c and vy = round(y) = Wcrm (x) = · −|x|3 + 5 · |x|2 − 8 · |x| + 4 for 1 ≤ |x| < 2 ,
2
by + 0.5c, b·c stands as the floor function. In the same way 0 for|x| ≥ 2

as in a 1D signal, the 2D interpolation may be described as (7)
a linear convolution operation (linear filter). The 2D kernel to 
3 · |y|3 − 5 · |y|2 + 1 for 0 ≤ |y| < 1
implement the Nearest-Neighbor interpolation [10], [17] can 1
be computed by : Wcrm (y) = · −|y|3 + 5 · |y|2 − 8 · |y| + 4 for 1 ≤ |y| < 2 ,
2
0 for|y| ≥ 2
( 
1 for − 0.5 ≤ x, y < 0.5 (8)
Wnn (x, y) = , (2)
0 otherwise The benefit of being able to apply the separation of variables
in the Bicubic interpolation method, as well as on its deriva-
and its graphic representation is depicted in Fig.1(a). tions, allow to work with both expressions 7 and 8, so that an
B. Bilinear extension of the Catmull-Rom interpolation method from 1D
to 2D can be made [20], for this it must be considered:
The 2D extension of the Linear interpolation method is
so-called Bilinear [19]. Expressed as a linear combination Wcrm (x, y) = Wcrm (x) · Wcrm (y), (9)
convolution filter, the 2D kernel [18] Wbil (x, y) can be stated
as the product of two kernels in 1D Wlin (x) and Wlin (y), The 2D kernel of this interpolation method can be visualized
this is: in the Fig.1(d).

Wbil (x, y) = Wlin (x) · Wlin (y) E. Lanczos


(
1−x−y+x·y for 0 ≤ |x|, |y| < 1 , (3) The Lanczos interpolation algorithm belongs to the family
= of methods based on the Sinc(·) function [18]. Unlike the
0 otherwise
methods described in the previous sections, they do not use
In Fig.1(b) it is depicted the 2D kernel of the Bilinear a polynomial (or other) approach to the function Sinc(·),
interpolation. but the function Sinc(·) in itself combined with an adequate

978-1-7281-1145-2/19/$31.00 ©2019 IEEE 113


(a) Nearest-Neighbor (b) Bilinear (c) Bicubic

(d) Catmull-Rom (e) Lanczos

Fig. 1. 2D interpolation kernels.

window function ψ(x), therefore an interpolation kernel takes flexibility in the implementation, and in addition reducing the
the following form: arithmetic complexity and bandwidth usage of the computation
for each data point. How it was stated in Section II, the
w(x) = ψ(x) · Sinc(x), (10)
interpolation algorithms can be implemented via separable
w(y) = ψ(y) · Sinc(y), (11) convolution filters; therefore, this processing mode is ideal to
be implemented in a heterogeneous system.
The particular window functions for the interpolation of
Lanczos are defined as:
 IV. H ETEROGENEOUS IMPLEMENTATION
1
 for|x| = 0
ψLn (x) = sin(πx/n)
πx/n 0 < |x| < n , (12) Graphics Processing Units (GPUs) are devices with high
 performance that have been widely used to accelerate image
0 for|x| ≥ n

processing tasks, or complex algorithms [21]. Introduced by

NVIDIA, the Compute Unified Device Architecture (CUDA)
1
 for|y| = 0
ψLn (y) = sin(πy/n)
0 < |y| < n , (13) enables GPUs to be devices for parallel computation. Nev-
πy/n
 ertheless, the graphics accelerators can be considered as co-
0 for|y| ≥ n

processors that receive instructions as well as tasks from the
where n ∈ N denotes the order of the filter. In the literature, CPUs, in this regard, together both devices form a hetero-
values of n = 2, 3 are suggested for digital image processing geneous system. In a conventional way, the CPU invokes a
tasks. From these separable 1D kernels in both x as well as kernel function, which encapsulates a set of instructions (in
y, it is possible to make the extension to 2D signals in the this research paper it can be the 2D interpolation kernels)
following way: executed simultaneously by multiple threads in the GPU;
then information processed is returned to CPU. In order to
WLn (x, y) = wLn (x) · wLn (y), (14) develop the interpolation task, a brain MRI image I(x, y)
In Fig.1(e) is depicted the 2D kernel of this interpolation with size M × N is transfered from the system memory
method. to the device memory, by using the cudaM emcpy function
with the macro cudaM emcpyHostT oDevice, then the pa-
III. I NTERPOLATION VIA SEPARABLE CONVOLUTION rameters threadsP erBlock = (32, 32, 1), numBlocks =
FILTERS (M/dimBlock.x, N/dimBlock.y, 1) and the texture are set
In conventional way, a 2D convolutional filter with a kernel up; after that, the kernel function is invoked by the CPU.
with width n and height m demands n × m multiplications Inside of kernel function, the selected interpolation technique
for each output pixel. Separable filters are a special type of is developed using the concept of separable convolution by
filter that can be expressed as the composition of two 1D the computation into horizontal (row) and vertical (column)
filters, one on the rows on the image, and one on the columns. passes; in other words, row and column filters are applied
A separable filter can be divided into two consecutive 1D onto the input image. The resulting image I(x, ˜ y) is copied
convolution operations on the data, and therefore requires only back to the CPU via cudaM emcpy function with the macro
n+m multiplications for each output pixel [26]. Convolutions cudaM emcpyDeviceT oHost; finally, the restored image is
based on separable filters have the benefit of offering more saved at hard disk. As it is stated in the CUDA C Programming

978-1-7281-1145-2/19/$31.00 ©2019 IEEE 114


Guide [22], the threads may access data from multiple memory B. BraTS 2017
spaces during their execution.
The performance of the interpolation algorithms is evaluated
by using the real database BraTS 2017. This consists of 285
V. R ESULTS
complete studies in the training data and 47 in the validation
The performance and the execution time of the evaluated data. In turn, each study has information of magnetic reso-
algorithms are quantified and compared through images ob- nance images acquired under the modalities T 1, T 2, T 1ce,
tained from Multimodal Brain Tumor Segmentation Challenge F lair and ground truth (for tumor detection). It is convenient
(BraTS) 2017 [23]. The experimentation was developed in to highlight that the T 1ce and F lair modalities contain more
a heterogeneous system consisting of an Intel(R) Core(TM) information to identify the abnormalities (tumors) present in
processor i7-4720HQ CPU @2.60GHz with 4 cores and 16 the studies; in this regard, increasing the resolution in these
GB of RAM memory; as well as, a GPU GTX Titan X, with images through interpolation methods could help to delimit the
24 multiprocessors and 3072 cores and 12 GB RAM memory regions where abnormalities exist. For this experiment, only
dedicated. The project was programmed in C language in the folders T 1, T 2, T 1ce and F lair were considered, each one
addition to NVIDIA Toolkit CUDA 9.1. consists of 155 images, therefore each study has 640 images
with size of 240 × 240 pixels, in JPG format with dynamic
A. Metrics range of 8 bits. In order to work with these images it was
necessary to convert them to PGM format, since CUDA in its
To objectively evaluate the quality of the resolution increase
conventional form only accepts images in PGM (gray scale)
in magnetic resonance images, four aspects are considered.
and PPM (color) formats.
The first is related to the performance of the algorithm and is
evaluated using the Peak Signal-Noise Ratio (PSNR) [24]. The
second is to quantify the preservation of the fine details of the TABLE I
AVERAGE EVALUATION MADE TO B RATS 2017 STUDIES .
image, it is determined by the Mean Absolute Error (MAE)
[24]. The third is the Structural Similarity Index (SSIM) [25], Modality Algorithm PSNR (dB) MAE SSIM RT (s)
N. Neighbor 34.9580 1.1506 0.9353 5.0346
which is a quantitative measure of the difference between the Bilinear 36.4109 1.0639 0.9773 2.0448
original and reconstructed image in terms of its luminances, T1 Bicubic 37.3699 0.9758 0.9864 4.8649
contrasts and structure information. These three metrics can Catmull-Rom 36.2633 0.9821 0.9768 1.7586
Lanczos 36.3687 0.9799 0.9764 4.2567
be determined from the following expressions:
N. Neighbor 34.5342 1.1642 0.9127 4.9458
Bilinear 35.0840 1.0860 0.9685 2.1879
(max(x(i, j)))2
 
P SN R = 10 · log , (15) T2 Bicubic 35.0569 1.0896 0.9586 4.9405
M SE Catmull-Rom 35.6104 1.0835 0.9694 2.1270
Lanczos 35.0556 1.0963 0.9685 4.1652
where MSE is the Mean Quadratic Error and is determined N. Neighbor 34.6892 1.1578 0.9683 5.9946
Bilinear 36.9781 0.9877 0.9612 4.6353
by: T 1ce Bicubic 39.6939 0.8269 0.9781 5.2055
Catmull-Rom 37.8181 0.8567 0.8915 2.7360
M X
N
1 X Lanczos 39.1822 0.8361 0.9671 4.1241
M SE = [x(i, j) − ê(i, j)]2 , (16) N. Neighbor 32.3135 1.2108 0.9480 5.2392
M ·N i=1 j=1 Bilinear 34.8505 1.0439 0.9243 3.8528
F lair Bicubic 37.4409 0.8993 0.9582 4.4307
where M · N represents the size of the images that are Catmull-Rom 35.4685 0.9255 0.8770 1.9470
Lanczos 36.7367 0.9067 0.9281 4.3383
being analyzed, x(i, j) is the original image and ê(i, j) is the
improved image. On the other hand, MAE can be computed
An average quantitative summary of each method and
by:
metric considered in this experiment is summarized in Table
1 XX
M N I. For this calculation, all 285 studies of the training folder
M AE = [x(i, j) − ê(i, j)], (17) were considered, so independently for each acquisition mode
M · N i=1 j=1
T 1, T 2, T 1ce and Flair were considered 44 175 images . A
pertinent clarification that must be made is that the images
The SSIM metric in a simplified form is calculated by the
were processed in their original condition; in other words,
following expression:
no pre-processing stage was applied to them. From these
(2µx µy + C1 ) · (2σxy + C2 ) results, it is notorious to observe that the Bicubic and Lanczos
SSIM (x, y) = , (18) interpolation methods have the best performance, as suggested
(µ2x µ2y + C1 ) · (σx2 + σy2 + C2 )
by the three metrics to evaluate the quality of the interpolation.
where x is the original image, y is the refined image, µx and However, the same can not be said in relation to processing
µy the luminance values, σx and σy are the contrast values, time, in this aspect Catmull-Rom method was the fastest, and
C1 and C2 are two constant values. Finally, the Runtime was its performance is very close to the Bicubic interpolation,
last aspect considered in the experimentation, with units of due to this capacity in the literature is suggested as a good
seconds. alternative, since it has a compromise between quality and

978-1-7281-1145-2/19/$31.00 ©2019 IEEE 115


runtime. In graphic form, these results can be visualized in the
Figure 2. To exemplify the results on real images, a sample
image was taken at random, in its four modalities, these and
the results of the interpolation are presented in the Figure 3.
The complexity of this database is exposed in these images,
since the 4 normal regions that exist (gray matter, white
and cerebrospinal fluid, background), as well as the present
abnormalities have the problem of variation of information;
that is to say, the contours of these regions are poorly defined,
which influences that they can not be delimited. Even so, the
interpolation methods were able to increase the resolution of
the images, which for us translates into a better delimitation of
the borders or borders of the different regions in the magnetic
(a) PSNR (dB)
resonance images.
VI. C ONCLUSIONS
In this research work, five accelerated interpolation methods
were evaluated. The results obtained suggest that the Bicubic
method obtained the best performance, however with a high
execution time. For its part, the Catmull-Rom method obtained
a balance between quality and processing time, in addition is
able to increase the sharpness of the images. The implemen-
tation in a heterogeneous system (CPU+GPU) is of great help
in the processing in real time, and offers the possibility of
processing batches of images or complete studies of magnetic
resonance images, as was exemplified in the experimental part.
As future work, this work can be extended in order to be able (b) MAE
to use as super-resolution algorithms.
ACKNOWLEDGMENT
The authors thank to CONACYT as well as
TecNM/CENIDET for their financial support through
the project 5688.16-P named Sistema para procesamiento de
imágenes de resonancia magnética para segmentación 3D y
visualización de tejidos cerebrales.
R EFERENCES
[1] Saisho, H., & Yamaguchi, T. (2004). Diagnostic imaging for pancre-
atic cancer: computed tomography, magnetic resonance imaging, and
positron emission tomography. Pancreas, 28(3), 273-278.
[2] Kalaiselvi, T., Sriramakrishnan, P., & Somasundaram, K. (2017). Survey
of using GPU CUDA programming model in medical image analysis. (c) SSIM
Informatics in Medicine Unlocked, 9, 133-144.
[3] Laal, M. (2013). Innovation process in medical imaging. Procedia-Social
and Behavioral Sciences, 81, 60-64.
[4] Ali, M., Khan, M., & Tung, N. T. (2018). Segmentation of dental X-
ray images in medical imaging using neutrosophic orthogonal matrices.
Expert Systems with Applications, 91, 434-441.
[5] Brown, G., Radcliffe, A. G., Newcombe, R. G., Dallimore, N. S.,
Bourne, M. W., & Williams, G. T. (2003). Preoperative assessment
of prognostic factors in rectal cancer using highresolution magnetic
resonance imaging. British Journal of Surgery, 90(3), 355-364.
[6] Saslow, D., Boetes, C., Burke, W., Harms, S., Leach, M. O., Lehman,
C. D., ... & Smith, R. A. (2007). American Cancer Society guidelines
for breast screening with MRI as an adjunct to mammography. CA: a
cancer journal for clinicians, 57(2), 75-89.
[7] Singh, J. F., & Magudeeswaran, V. (2017, February). Thresholding
based method for segmentation of MRI brain images. In I-SMAC (IoT
in Social, Mobile, Analytics and Cloud)(I-SMAC), 2017 International
Conference on (pp. 280-283). IEEE. (d) Time (s)
[8] Luo, J., Mou, Z., Qin, B., Li, W., Yang, F., Robini, M., & Zhu, Y. (2017).
Fast single image super-resolution using estimated low-frequency k- Fig. 2. Graphic trend of the average evaluation made to BraTS 2017 studies.
space data in MRI. Magnetic resonance imaging, 40, 1-11.

978-1-7281-1145-2/19/$31.00 ©2019 IEEE 116


(a) T 1 (b) Nearest Neighbor (c) Bilinear (d) Bicubic (e) Catmull-Rom (f) Lanczos

(g) T 2 (h) Nearest Neighbor (i) Bilinear (j) Bicubic (k) Catmull-Rom (l) Lanczos

(m) T 1ce (n) Nearest Neighbor (o) Bilinear (p) Bicubic (q) Catmull-Rom (r) Lanczos

(s) F lair (t) Nearest Neighbor (u) Bilinear (v) Bicubic (w) Catmull-Rom (x) Lanczos

Fig. 3. Sample image and its results after the interpolations.

[9] Qifang, X., Guoqing, Y., & Pin, L. (2017). Super-resolution Reconstruc- application to human cardiac DT-MRI. Medical image analysis, 16(2),
tion of Satellite Video Images Based on Interpolation Method. Procedia 459-481.
Computer Science, 107, 454-459. [18] Burger, W., & Burge, M. J. (2016). Digital image processing: an
[10] Lehmann, T. M., Gonner, C., & Spitzer, K. (1999). Survey: Interpolation algorithmic introduction using Java. Springer.
methods in medical image processing. IEEE transactions on medical [19] Li, X., & Orchard, M. T. (2001). New edge-directed interpolation. IEEE
imaging, 18(11), 1049-1075. transactions on image processing, 10(10), 1521-1527.
[11] Izadpanahi, S., & Demirel, H. (2013). Motion based video super reso- [20] Wu, T., Bai, B., & Wang, P. (2013). Parallel catmull-rom spline
lution using edge directed interpolation and complex wavelet transform. interpolation algorithm for image zooming based on CUDA. Applied
Signal Processing, 93(7), 2076-2086. Mathematics & Information Sciences, 7(2), 533.
[12] Qifang, X., Guoqing, Y., & Pin, L. (2017). Super-resolution Reconstruc- [21] Snchez, M. G., Vidal, V., Bataller, J., & Arnal, J. (2013). A parallel
tion of Satellite Video Images Based on Interpolation Method. Procedia method for impulsive image noise removal on hybrid CPU/GPU systems.
Computer Science, 107, 454-459. Procedia Computer Science, 18, 2504-2507.
[13] Yang, Q., Zhang, Y., Zhao, T., & Chen, Y. (2017). Single image [22] Nvidia, C. U. D. A. (2018). Programming guide.
super-resolution using self-optimizing mask via fractional-order gradient [23] Menze, B. H., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani,
interpolation and reconstruction. ISA transactions. K., Kirby, J., ... & Lanczi, L. (2015). The multimodal brain tumor
[14] Zhang, Q., & Wu, J. (2015). Image super-resolution using windowed image segmentation benchmark (BRATS). IEEE transactions on medical
ordinary Kriging interpolation. Optics Communications, 336, 140-145. imaging, 34(10), 1993-2024.
[15] Alaya, I. B., Jribi, M., Ghorbel, F., Sappey-Marinier, D., & Kraiem, T. [24] Mújica-Vargas, D., de Jesús Rubio, J., Kinani, J. M. V., & Gallegos-
(2017). Fast and Accurate Estimation of the HARDI Signal in Diffusion Funes, F. J. (2018). An efficient nonlinear approach for removing fixed-
MRI Using a Nearest-Neighbor Interpolation Approach. IRBM, 38(3), value impulse noise from grayscale images. Journal of Real-Time Image
156-166. Processing, 14(3), 617-633.
[16] Neubert, A., Salvado, O., Acosta, O., Bourgeat, P., & Fripp, J. (2012). [25] Wang, Z., Simoncelli, E. P., & Bovik, A. C. (2003, November). Mul-
Constrained reverse diffusion for thick slice interpolation of 3D volumet- tiscale structural similarity for image quality assessment. In Signals,
ric MRI images. Computerized Medical Imaging and Graphics, 36(2), Systems and Computers, 2004. Conference Record of the Thirty-Seventh
130-138. Asilomar Conference on (Vol. 2, pp. 1398-1402). Ieee.
[17] Yang, F., Zhu, Y. M., Magnin, I. E., Luo, J. H., Croisille, P., & Kingsley, [26] Podlozhnyuk, V. (2007). Image convolution with CUDA. NVIDIA
P. B. (2012). Feature-based interpolation of diffusion tensor fields and Corporation white paper, June, 2097(3).

978-1-7281-1145-2/19/$31.00 ©2019 IEEE 117

View publication stats

You might also like