The new system has been applied to a variety of MRT test sequences and yields a good noise and artifact reduction which is evaluated by Signal-to-Noise-Ratio.
Image quality enhancement for MRT images Holger Blume, Stephan Kannengiesser and Tobias G. Noll Chair for Electrical Engineering and Computer Systems RWTH Aachen, Schinkelstraße 2, 52062 Aachen, Germany Phone: +49 (0)241 80 97591 Fax: +49 (0)241 80 92282 E-mail: {blume|kan|tgn}@eecs.rwth-aachen.de
Abstract— In this contribution a new combination of methods for noise and artifact reduction is presented. The new noise reduction system has successfully been applied to magnetic resonace tomography (MRT) sequences. For this purpose a system which is based on a nonlinear bandsplitting has been elaborated and optimized. In order to cope with motion artifacts and with ghosting effects, the motion detection is performed in only one of the resulting nonlinear subbands where noise, ghosting effects and objects can be effectively differentiated. This motion information is applied then to adjust the coefficients of subsequent recursive filtering processes. Furthermore, a coring is applied in each subband. The new system has been applied to a variety of MRT test sequences and yields a good noise and artifact reduction which is evaluated by Signal-to-Noise-Ratio (SNR) gains over a wide range of input SNRs. In this application, the new approach outperforms classical noise reduction techniques and achieves Peak-Signal-to-Noise Ratio gains of up to 7 dB for typical input sequences. Additionally, an implementation study has been performed in order to compare different implementation alternatives for this new method.
price systems with reduced field strength are coming up. These systems may be applied in future for applications in the medical field (e.g. medical testing for extremities like Achilles tendon) or for technical inspections like non-destructive material testing. Compared to high end systems the image quality of such low field systems is much lower in terms of the achievable Signal-to-Noise Ratio (SNR). One example for a prototype of a dedicated system is the MRT surface scanner which has been developed in a joint project at RWTH Aachen. In future, this system will facilitate mobile applications of MRT [8]. Figure 1 depicts the first prototype of this mobile MRT surface scanner.
Keywords— magnetic resonance tomography (MRT), image quality enhancement, motion adaptive noise reduction, nonlinear filter bank
I.
INTRODUCTION
Magnetic resonance tomography (MRT) is a very powerful imaging method in medical diagnostics. It features excellent soft tissue contrast and low patient exposure (e.g. no x-rays are applied). In this field, two development trends can be observed: On one hand, besides static image sequences, dynamic real time sequences are becoming more and more important. Examples for this are functional studies at the heart muscle or at moving wrists or ankles. On the other hand, besides classical high field systems, small, dedicated low
Figure 1 Prototype of a Mobile System for Unilateral MR Tomography, the MRT Surface Scanner [8]
Featuring a magnetic field of 0,2 T this system lies in the upper range of low field systems. Due to the compact set-up and the resulting reduced sensitive volume (Field of View, FOV) a drastically reduced SNR has to be accepted. However, in order to enable dynamic studies like functional studies of an Achilles tendon, the resulting SNR has to be improved by intensive postprocessing of the image material. In principle this is
a known problem from video signal processing, but in the context of MRT the key parameters are different in the sense of a reduced frame rate, a lower image resolution and especially a much lower SNR. Furthermore, MRT specific artifacts arise. The reason for these MRT specific artifacts lies in the applied methods for data acquisition. The duration of a complete acquisition of the spatial frequency spectrum is usually too long in order to "freeze" a motion process. Therefore, artifacts like local blurring (unsharpness) or even shifted repetitions of the objects within an image (ghosting) arise. A noise and artifact reduction based on the analysis of image series has to provide at first instance robustness against these characteristic artifacts. In this contribution a new combination and adaptation of known methods for noise and artifact reduction to MRT sequences is described. For this purpose a system which is based on a nonlinear band splitting according to [6] has been elaborated and optimized. In order to cope with motion artifacts, the motion detection is performed in only one of the resulting nonlinear subbands, i.e. it is performed in the coarse granular subband. Only in this subband noise, ghosting effects and objects can be effectively differentiated. This motion information is applied then to adjust the coefficients of the subsequent recursive filtering processes. Furthermore, a coring is applied in each subband. This new approach has been applied to a variety of MRT test sequences (simulated and measured images from high and low field set-ups) and yields a good noise and artifact reduction which can be expressed e.g. by SNR gains over a wide range of input SNRs. In this application, the new approach outperforms classical noise reduction techniques. An implementation study has been performed in order to compare different implementation alternatives for this new method. Here, the achievable computation times and the associated power figures for different implementation alternatives (FPGA based, DSP based, general-purpose-processor based) are shown. The paper is organized as follows: In chapter II the applied methods for image quality enhancement are discussed. Here, especially the principle of nonlinear band splitting in combination with a motion adaptive noise reduction is explained and the adaptation of this method to the MRT images is worked out. The simulation results for the new noise and artifact reduction system are given. A second subsection of this chapter explains a further nonlinear synthetic edge enhancement technique which is applied to the noise reduced images in order to increase the image sharpness. In chapter III the results of
an implementation study, comparing different implementation alternatives are discussed. The results of the paper are summarized in chapter IV. II.
IMAGE QUALITY ENHANCEMENT FOR MRT IMAGES
A. State-of-the-art techniques For many applications the (subjective) image quality is of great importance. Therefore, the reduction of artifacts and the reduction of noise is a key challenge. Many of the techniques which are applied today for the removal of noise are based on an adaptive application of linear spatial operators (see e.g. [2], [7], [10]). But this principle has some drawbacks especially for edges. Therefore, the performance of these methods is limited. Furthermore, there are some methods which are based on a subband splitting and a succeeding processing of the separate subbands (e.g. [5]). In this context different methods for subband splitting such as linear as well as pyramidal subband splitting have been presented. But also for subband splitting it has been shown that again linear methods possess some drawbacks in the presence of sharp edges and therefore cannot be used efficiently for noise reduction in images. One attractive alternative which has been studied recently is nonlinear band splitting. Generally, the goal of band splitting techniques is the extraction of image elements showing specific features. The application of linear filters for the band splitting leads to subbands which differ from each other concerning their frequency content. But in case of a band splitting for noise reduction this kind of signal decomposition can be hindering. Typical image content is composed of a variety of edges which consist of different frequency components. Therefore, signal components are submitted to several subbands and as a consequence of this the differentiation between noise and signal components becomes extremely complicated. Consequently, around edges often either no noise reduction is performed or signal components of an edge are recognized as noise and therefore are removed from the image. This leads to a degradation of the image quality. Therefore, methods are required, which allow to submit edges as a whole to one subband and not to distribute them over several subbands. This can be achieved by application of nonlinear subband splitting. For instance in [9] a first nonlinear band splitting applying median filters has been presented. By application of N different median filters the signal has been decomposed in N+1 different subbands. By this
decomposition technique different subbands arise whose signals correspond to image elements of different sizes. One drawback of this technique is that the size of the filter masks is increasing drastically with the number of subbands (e.g. exponential increase). This results in a very high computational effort. In [6] this drawback has been overcome by application of cascaded median filters. The goal of this cascading is the separation of signal components with an increasing size from subband to subband. Applying a nonlinear filter cascade it is also possible to reduce the computational effort as the single median filters apply suitable weighting masks which allow to use only a small number of non-zero weighting coefficients. By application of this kind of efficient nonlinear band splitting it could be shown in [6] that also in case of strongly corrupted noisy images the edge slopes remain nearly unchanged and the separation of image elements within single subbands is facilitated. Therefore, a simple differentiation between image elements and noise becomes possible due to the different signal amplitudes. In [6] the final noise reduction is performed by application of a soft-coring technique to the single subbands. A benchmarking of this technique shows that it is superior to other state-of-the-art techniques in noise reduction. This method has been the basis for the noise reduction which has been developed here for strongly corrupted MRT images. B. Nonlinear band splitting and motion adaptive noise reduction The goal of this work was to provide noise reduction techniques which can be applied also for image sequences with moving objects. Classically this could be solved by application of motion adaptive filter techniques which adapt the filter coefficients for the noise reduction filter according to the detected degree of motion [3] (i.e. no noise reduction for fast moving objects in order to avoid motion smear/blurring, respectively full noise reduction in case of non-moving image elements). The problem of the MRT image sequences discussed here is that the available motion detectors often fail as the image quality of the mobile MRT surface scanner is rather low due to several reasons (see chapter I). There-
fore, image artifacts (e.g. double contours) are often wrongly detected as motion whereas really moving objects are not. The key to the solution of this problem is the application of the nonlinear band splitting. The separation of different object sizes allows to perform the motion detection only in the coarse granular signal subband. Here, only the coarse granular image elements are included. They are separated in this channel totally from noise or artifacts like double contours. If the motion detection is performed here, a reliable motion information (at least for the coarse granular image elements) is achieved. This reliable motion information can be applied in a classical way for adapting the filter coefficients of recursive IIR (infinite impulse response) filters for noise reduction. Besides this recursive filtering also an additional soft-coring can be performed in each subband. The resulting complete system is depicted in the block diagram in Figure 2. Here, also the applied cascaded weighted median filters are depicted. It is visible that four different median filters are applied which lead to five different subbands. The cascading allows to apply only five non-zero coefficients for each median filter. The motion detection is performed in the coarse granular subband (subband 5). In Figure 3 the nonlinear band splitting is depicted for an exemplary MRT image of a high field MRT scanner. It is visible that from band to band more coarse granular elements are included in the corresponding subband images. In order to evaluate the performance of this system several test sequences (simulated and measured images from high and low field set-ups) have been applied. For this purpose, the simulated or measured test data have been subject of an addition of typical artifacts respectively an addition of noise. Then these test sequences were applied for the evaluation of the new noise reduction technique. As a quantitative metric for image quality assessment the PSNR (Peak Signal to Noise Ratio, see for example [4]) has been applied.
Image with artifacts
256
256
Band 1
IIR 256
256
Band 2
Med1 Med1
IIR 256
1 131 1
256
Band 3
Med2 Med2
IIR 256
1 111 1
256
Band 4
Med3 Med3
1 0 10101 0 1
Med4 Med4
1 0 0 1001001 0 0 1
IIR
Band 5
Filtered image
Motion Motion detection detection
Figure 2 Nonlinear Filterbank and motion adaptive noise and artifact reduction
Original image
Band 1
Band 2
Band 3
Band 4
Band 5
Figure 3 Nonlinear band splitting applied to an exemplary MRT image
The PSNR is a standard criterion for objective evaluation of image quality PSNR[dB ] = 10 ⋅ log 1 ⋅ N ⋅ M
N
M
255 2 2 F pic [i, j ] − Fref [i, j ]
∑∑ ( i =1 j =1
.
)
Here, N, M denote the picture height and width, Fpic[i,j] denotes the pixel value at i,j and Fref[i,j] the pixel
value of an ideal reference picture (a resolution of eight bit per pixel is assumed here). The PSNR therefore yields the relation between the maximum signal power and the power of the distortion error. In Figure 4 a) four images out of a synthetic phantom sequence featuring moving objects are depicted. A simulated corruption which produces MRT specific artifacts and additive white Gaussian noise has been applied to this sequence (Figure 4 b).
t=10
t=20
t=30
t=40
a)
b)
c) Figure 4 Four images out of a sequence of forty images for a) synthetic phantom images featuring moving objects; b) simulated corrupted synthetic phantom images (PSNR 28,4 dB); c) postprocessed synthetic phantom images
For example the small ball-like structure in the upper right corner of the phantom image which is moving with a velocity of 1 pixel/frame in horizontal and vertical direction is extremely blurred in the simulated corrupted image. Furthermore, ghosting effects of this ball due to reconstruction errors are visible. The ring in the upper part of the image is zoomed from frame to frame with a zoom velocity of approximately -0,2 pixel/frame. This zoom and the reconstruction errors lead to echo contours in the simulated corrupted images. The resulting postprocessed images in Figure 4 c) show that many of the annoying artifacts could be reduced. For example the echo contours of the zooming ring could be totally removed. Furthermore, the ghosting effects around the moving ball can be reduced. Finally, the additive Gaussian noise is also drastically reduced. In Figure 5 results for the noise reduction capability of
the new method applied to the synthetic phantom images featuring moving objects are depicted. The input PSNR has been varied from 32 to approximately 24 dB. It is noticeable that the noise reduction gain of the new method increases with degrading input PSNR. For example for an input PSNR of 32 dB a noise reduction gain of 1.5 dB can be achieved whereas for an input PSNR of 24.3 dB a noise reduction gain of 7 dB can be achieved. By this nearly a constant image quality over a wide range of input PSNR can be achieved. The new method has been compared also to classical noise reduction techniques like motion adaptive IIR filtering (without linear band splitting) and a band splitting technique in combination with static coring. The resulting curves in Figure 5 prove that these basic techniques achieve noise reduction gains of about 2 dB at maximum. Only the combination of nonlinear band splitting and motion adaptive filtering and coring is able
to yield a robust and significant quality gain.
8 7
PSNR gain [dB]
8
PSNR gain [dB]
7 6 5 4 3 2
5 4 3 2 1 0
1
33
0 32,0
31,3
30,7
30,1
28,4
27,2
26,2
25,6
7 subbands
adapt. IIR-filter
Coring
The same analysis has been performed for measured images from high field MRT devices. Also for this data a simulated corruption has been applied in order to simulate artifacts of a low field device. In Figure 6 the results for an MRT image of a head is depicted. Also for this image significant noise reduction gains up to 11 dB could be achieved by application of the new method. 12 10 8 6 4 2 0 33,7
30,5
28,7
27,4
26,6
input PSNR [dB]
Band splitting
adapt. IIR-filter
31
30
29
28
27
26
25
24
23
input PSNR [dB]
Figure 5 Results for synthetic phantom images featuring moving objects
35,0
32
24,3
input PSNR [dB]
Band splitting
PSNR gain [dB]
6
Coring
Figure 6 Results for the test sequence "Head"
C. Influence of the applied number of subbands It has been explored how many subbands should be applied for an efficient noise reduction. Therefore, several test images have been processed applying the new method and different numbers of subbands. Figure 7 shows the results for the synthetic phantom images which have been introduced before. It is obvious that the achievable quality gain depends on the number of applied subbands. A drastic quality increase can be achieved by applying more than three or four subbands. Then the quality gain saturates. The number of subbands also has a strong effect on the required computational power (see chapter III). Therefore, the number of subbands has been fixed here to five.
6 subbands
5 subbands
4 subbands
3 subbands
Figure 7 Influence of the number of subbands (synthetic phantom images)
After the prototype of the MRT Surface Scanner had been finished it had been applied for picking up test images. One of the first test images has been a letter ("M") which was manufactured of rubber. In Figure 8 three exemplary images (Nr. 20 - 22) out of a sequence of 64 test images which were picked up in sequential order (no movement of the test object, image size 128x128 pixel, image pick-up time per image approximately 20 s) are depicted. The reduction of the MRT specific distortions are clearly visible. Whereas in the single original images the bars of the letter "M" are disrupted they are much more smooth and clear in the postprocessed images.
Nr. 20
Nr. 21
Nr. 22
Figure 8 Test image rubber letter; first row original images with MRT specific distortions, second row postprocessed images
D. Nonlinear edge enhancement One remaining drawback of the postprocessing is a reduction of image sharpness. In order to overcome this problem a nonlinear synthetic edge enhancement technique according to [11] has been applied to the filtered output sequence. First, within the filtered output sequence an edge detection is performed. Here, a Sobeloperator (see e.g. [4]) has been applied as it features a noise robust behavior. Then, according to the extension and to the height of the current edge which has been detected, a correction signal which is stored in an adjustment table is selected. This synthetic detail signal is added to the original (unsharp) signal. Figure 9 depicts the block diagram according to [11] and Figure 10 the principle of the edge enhancement process. enhanced signal senh (x)
original signal s(x)
adjustment signal d(x) edge height/extension Edge detection and parameter calculation
-s
s
ag
position flag
Adjustment table
Figure 9 Block diagram for the synthetic edge enhancement [11]
s(x)
(a)
edge extension edge height
Before Edge enhancement
x
d(x) (b)
Adjustment signal
x
senh(x) (c)
Edge enhancement
x Figure 10 Principle of synthetic detail signal addition, a) before the edge enhancement, b) synthetic detail signal, c) after edge enhancement through addition of detail signal
This edge enhancing technique has also been applied to the filtered output sequence which was picked-up with the MRT prototype. Figure 11 a) shows a single image
from the test sequence discussed before and the result of an edge detection by application of a horizontal Sobeloperator. Figure 11 c) depicts the final output image after the edge enhancement process (applied separately for horizontal, vertical and diagonal edges). It is visible that the sharpness of the image could be increased by this method.
a)
b)
c) Figure 11 Edge Enhancement applied to a test image which was picked up with the mobile MRT surface scanner a) Original image after noise and artifact reduction b) result of horizontal Sobel-operator applied to that image c) result of nonlinear, synthetic edge enhancement
III.
IMPLEMENTATION STUDY
The main feature of the mobile system for MR tomography is its mobility. Therefore, the required postprocessing for the output images has also to be provided by mobile computation devices. It is possible to apply mobile computers (laptops) for this task or to equip these computers with some coprocessing devices (e.g. coprocessor boards) which are able to perform the required computations in an acceptable time and featuring acceptable low power consumption. Hence, some implementation alternatives ranging from general purpose processor implementations to digital signal processor implementations or FPGA based implementations have been evaluated in terms of required hardware effort. Here, especially the required computation time and the required power consumption are key features.
In Table 1 the resulting features of these implementation alternatives are listed. Whilst the computation time and power consumption are given quantitatively the flexibility is rated qualitatively. Flexibility here means the ability to exchange a given algorithm setting which could be measured e.g. in the reciprocal of the reprogramming resp. reconfiguration time. A general purpose or digital signal processor can be programmed for example in a high level language and therefore the program can be exchanged easily. A hardware implementation in form of an FPGA requires some more time to write the new HDL code and to verify the functionality of the program. It can be summarized that even a very moderate processor is able to perform the postprocessing in an acceptable computation time. The application of a modern DSP architecture drastically reduces this computation time. Finally, also an FPGA implementation is possible with only a small effort of required logic elements. Hence, it has been proven that the new postprocessing system can be realized on the basis of different implementation alternatives with a low required hardware effort. hardware platform
implementation costs
general purpose processor SUN SPARC 5 @ 70MHZ
T ≈ 1,4 s per 128x128 image PD ≈ 6 W flexibility + + (C-program)
digital signal processor (DSP) Trimedia TM1300@166MHz
T ≈ 0,02 s per 128x128 image PD ≈ 3 W flexibility + + (C- program)
field programmable gate array (FPGA) APEX EP20K200
T ≈ 0,2 s per 128x128 image PD ≈ 0,4 W #logic elements ≈ 2100 flexibility ~ ~ (VHDL- program)
Table 1
Specific features of different implementation alternatives for subband based noise and artifact reduction
IV.
SUMMARY
An image enhancement system for magnetic resonance tomography (MRT) sequences has been presented in this paper. Main elements of this image enhancement system are a motion adaptive noise reduction which is based on a nonlinear subband splitting and a synthetic edge enhancement system. The performance of the image enhancement has been demonstrated in terms of Peak Signal to Noise Ratios for a comparison between reference images and noise reduced images for a variety of MRT test sequences (simulated and measured images from high and low field set-ups). The new system has also been applied
successfully to the first output images which have been picked up by the prototype of a mobile MRT scanner. It has been shown that this new combination of different noise reduction techniques achieves good results which are superior to the other inspected methods. As the mobile MRT device shall be applied in the field the required hardware effort for postprocessing the image data is also a key issue. Therefore, an implementation study has been performed. It has been shown that only a low hardware effort is required for this new system. Even a very moderate processor is able to perform the required computations in an acceptable time. REFERENCES [1] Altera, Excalibur Backgrounder, http://www.altera.com/ literature/wp/excal_bkgrnd.pdf, White Paper, June 00 [2] G. de Haan, T. G. Kwaaitsaal-Spassova, M. Larragy, O. A. Ojo, Memory integrated noise reduction IC for Television, IEEE Trans. on CE, Vol. 42, No. 2, May 1996, pp.175-181 [3] J. O. Drewery, R. Storey, N. EW. Tanton, Video Noise Reduction, BBC Research Department Report, 1984/7. [4] A. K. Jain, Fundamentals of Digital Image Processing, Prentice Hall, Englewood Cliffs, NJ, 1989. [5] K. Jostschulte, A. Amer, M. Schu, H. Schröder, A subband based spatio-temporal noise reduction technique for interlaced video signals, Proc. of the International Conference on Consumer Electronics, Los Angeles, USA, 1998. [6] K. Jostschulte: Subjektiv angepasste Signalaufspaltung mit nichtlinearen kaskadierten Filtern, Proc. of the ITG Fachtagung Multimedia 2001, Dortmund, 2001 (in German). [7] J. S. Lee, Digital Image Smoothing and the Sigma Filter, Computer Vision Graphics Image Processing, Vol. 24, 1993, pp.255-269. [8] H. Popella, J. Felder, M. Wenzel, S. Kannengiesser, G. Henneberger, B. Rembold, T. G. Noll, A Mobile System for Unilateral NMR Tomography: The NMR Surface Scanner, Proc. ISMRM 2002, pp. 838. [9] P. Salembier, M. Kunt, Size-sensitive multiresolution decomposition of images with rank order based filters, Signal Processing, Vol. 27, No. 2, 1992, pp.205-241. [10] M. A. Tekalp, Digital Video Processing, Prentice Hall, 1995. [11] X. Wu, Synthetische Kantenversteilerung zur Verbesserung der Bildschärfe. PhD Thesis Universität Dortmund, VDI Fortschrittsberichte, Reihe, Nr. 272, 1993 (in German).
