Problems in texture analysis with magnetic resonance imaging.

Since its introduction in the 1930s, magnetic resonance imaging (MRI) has become recognized as a powerful in vivo diagnostic tool. The objective of this article is to discuss developments in quantitative MRI - and particularly texture analysis - that maximize diagnostic information, A fundamental part of the work involves careful study of the optimal MRI data collection strategies for texture analysis. This is critical, because different centers may vary their measuring sequences and acquisition protocols for clinical reasons, and may be reluctant to vary these for texture investigation. Different measuring techniques, such as spin echo, gradient echo, and echo planar, and different measuring parameters produce totally different patterns in texture. Careful investigation of the dependence of all these variables using texture phantoms (test objects) will help understand how MRI image texture is formed from tissue structures. Therefore, it is essential to design and test reliable and accurate test objects for a detailed assessment of texture analysis methods in MRI, The main feature of these test objects is their ability to simulate tissue-like textures with tissue-like MR relaxation properties. Long-term stability is also vital, as is uniformity of the overall texture. Another aspect is to examine the test objects under a whole range of MRI measuring sequences and imaging conditions using different scanners to determine their stability and utility.

Magnetic resonance imaging (MRI) is one of the most, exciting imaging technologies for texture analysis: it offers the best soft, tissue contrast, which can be dramatically varied during imaging. Careful study of the dependence of texture parameters on MRI data collection strategy is essential for texture analysis in order to avoid artificial texture from the scanner. This is critical, since different centers may vary their measuring sequences and acquisition protocols for their clinical investigations.

The basic problem in quantitative MRI texture analysis is the large number of different measuring techniques and imaging parameters, which can be easily changed during a clinical examination. Thus, different techniques and imaging parameters produce totally different patterns in the texture parameters of the same tissues in clinical examinations with different sensitivity to artificial texture overlaid by the scanner. The main problem in texture analysis with MRI is to avoid this artificial texture and minimize its influence. The presented work was performed in the framework of a European research project COST (Cooperation in the Field of Scientific and Technical Research) Bll between 1998 and 2002 by institutions from 13 European countries, aimed at the development of quantitative methods for MRI texture analysis.[1] For further detail of texture analysis, parameters, and software, see the article by Materka in this volume[2] or references 3 to 7.

Material and methods

The complexity of this problem can be demonstrated by considering a typical measuring spin echo sequence as measured by a commercial whole-body imager. Various parameters can be easily changed during clinical investigation: image contrast is mainly defined by repetition time (TR) and spin echo time (TE); image resolution is defined by slice thickness (TH), field of view (FOV), and matrix size (MA), which also influence texture analysis. The parameters of k-space acquisition and reconstruction arc very important: k-space is the artificial space in which the raw MRI data are collected, and the image contrast, and texture is very sensitive to k-space strategies. Other parameters like coil setup and number of active coil segments are also responsible for signal and flip angle (α) variations in the image. Careful investigation of the dependence of all these variables will help understand how MRI image texture is formed in tissue structures. In our studies, MRI acquisition was performed in the standard head coil of a 1.5-T scanner (Siemens Vision, Erlangen, Germany).

Spin echo technique

One of the most, important measuring techniques in clinical diagnosis is the spin echo sequence, in which 90° and 180° radio frequency (RF) pulses produce the spin echo signal. In addition, gradients are used in x,y, and z. directions to localize the signal.[8] The advantages of this technique are reduced artifacts, clearly defined contrast, and common availability. The disadvantages arc the contrast dependency on RF pulse quality, and slice cross-talking, which is typical of a two-dimensional (2D) technique.

This imaging technique allows measurement of the three relevant MRI tissue parameters: spin density (ρ), spin-lattice relaxation time (T1), and spin-spin relaxation time

(T2), which are most responsible for tissue contrast and texture. According to the theoretical equation for the spin echo signal:[9]

S ≈ ρ · (1—e) · e [1]

in which S is the spin echo signal, the contrast p can be created by a long TR and short. TE, resulting in a flat image contrast and texture at high signal intensity (. T1 contrast can be created by short TR and short TE in spin echo imaging (Figure 1b). On the other hand, T2 contrast is created by long TR and long TE, mainly reflecting the water content of the tissue (Figure 1c). These three physical tissue parameters are described in reference 1 .The real physical properties of tissues may be obscured by artificial contrast, and texture from the scanner.

Slice profile

Slice profile is defined by the slice gradient and the shape of the RF pulse. Ideally, we would like to measure a rectangular slice, but due to technical reasons the real slice profile is Gaussian shaped. The consequence is that we have signal contributions from neighboring slices that influence the tissue texture. To minimize this effect, an interleaved slicing scheme is used in multislice 2D imaging.

k-space

Another aspect of artificial texture is connected to the kspace, which describes the strategy for raw data collection. The k-space contains the measured signal frequencies k and k, the so-called hologram from which the real MRI image can be calculated by a Fourier transform(. Some imaging techniques measure only every second line in the k-space to speed up the imaging sequence, which results in a reduction in the signal-tonoise ratio (SNR) by 1/√2 and aliasing artifacts, with consequences for image texture. Restriction to the center of the k-space with zero filling of the outer part results in the same SNR effect without aliasing.

RF excitation

Another important variable is the RF characteristic and sensitivity of the transmitting and/or receiving coil, which can produce a lot of artificial texture from the scanner. This is demonstrated in using hard image scaling, which shows a clear signal inhomogcneity due to nonideal RF pulses at the outer range of the phantom (ie, coil).

Another coil effect on image texture is produced with coil arrays, where the summarized image is a result of the combination of single coils, each of which contributes its own coil characteristics (eg, SNR, sensitivity, and RF excitation profile) to the summarized image. This means that image texture could slightly differ between the object center and the object boundary, where protons are close or far away from the center of the coil.

Gradient echo techniques

Significant effects on image contrast and texture are introduced by the imaging sequence itself, since the imaging signal can have a very complex dependence on the physical properties of the underlying tissue. One example is the socalled gradient echo technique like FLASH (fast low angle shot),[10] where the 90° and 180° RF pulses are replaced by a low-angle RF pulse with a bipolar gradient scheme resulting in a gradient, echo signal. This measuring technique can be used as fast, imaging 2D technique or as a real 3D imaging technique because of the compact timing of the sequence. On the other hand, the FLASH signal has a complex dependence on T, the local spin-spin relaxation time (T2*), and the flip angle a, according to:

S ≈ A · e-

with A = ρ · sinα · (1—e-)/(1—cosα · e-TR/T1) [2]

Thus, the different flip angle distributions produced by the coil characteristics result in different signals, and as a consequence in different image texture patterns as demonstrated in (.

A very complex signal and texture situation is present in so-called single shot imaging techniques like echo planar imaging (EPI),[11] where k-space is filled in one shot with multiple gradient echoes. This is achieved by a gradient scheme in which the upper corner of the k-space is reached by a single gradient pulse followed by a series of blips resulting in a rectangular movement through the kspace.This technique is very sensitive to local susceptibility artifacts, resulting in image distortions and strong T2* contrast dependence.

Some special imaging techniques like spiral imaging can produce a very complex pattern in the image texture, since this single shot technique moves on a spiral through the k-space, which can be achieved by oscillating gradients with a phase shift of 90° in the x and y directions. This technique requires data interpolation in k-space to bring the measured data onto a Cartesian coordinate system before .Fourier transform. This interpolation can produce spurious artifacts with the consequence that the image texture is dependent on k-space interpolation and image reconstruction. In addition, the problem of texture dependence on measuring technique is more complicated due to the large number of imaging sequences available on modern scanners, as illustrated in (

Results and discussion

SNR dependence

show the results of a FLASH experiment, in a normal volunteer for SNR dependence measurement of texture parameters. The measuring parameters of the FLASH experiment were: TR/TE/α = 2 ms/ 9 ms/30°; bandwidth (BW) = 195 Hz/pixel; MA = 512x512; FOV = 280 mm; TH = 2 mm; and acquisitions (AQ) = 1 to 324 resulting in an SNR = 1 to 18. Texture parameters (SNR, entropy 5x5, correlation 5x5) of white matter, gray matter, and noise are shown as a function of the number of acquisitions (=SNR2). Figure 6c demonstrates that no texture can be measured in white matter using standard image resolution (0.5x0.5x2 mm3) as described above, since the SNR of white matter has the same characteristics as noise. In contrast, the SNR of gray matter reaches a nearly constant value at about 16 acquisitions and no further improvement can be reached due to the true underlying texture of the tissue. The same observation holds for a typical parameter of microtexture, like entropy 5x5 (Figure 6d), while no dependence on SNR can be detected for a typical parameter of macrotexture, like correlation 5x5 (Figure 6e). Based on this observation, a sufficient SNR>4 is necessary to measure the real textural behavior of the human brain.[12,13]

Normalization

A texture test object (PSAG) was developed on the basis of polystyrene (PS) and agar solution (AG) to mimic texture properties artificially. PS spheres are available from the technological process of PS production. Two types of spheres were used for the phantom construction: randomly distributed spheres of diameter 0.2 to 3.15 mm; or mechanically separated spheres of diameter 0.8 to 1.25 mm, 1 .25 to 2 mm, or 2 to 3.15 mm. Polyethylene tubes of diameter 1.5 and 2.8 cm were filled with spheres and by a hot solution of 4% agar (free and doped with DyCl3). One milliliter of 0.1 % NaN3 was added per liter of agar for microbiological stability.[14]

A second texture test object containing foam at different. densities in Gd-DTPA solution was used to describe microtexture properties. Phantom tubes containing foams with coarse, middle, and fine density were constructed and filled with a Magnevist® (Schering, Berlin) solution at. a concentration of 1:4000. Problems with the foam phantoms are air bubbles, which create susceptibility artifacts in the images, and so a careful preparation of the foam phantoms is necessary. Both types of phantoms were placed next to the head of a volunteer and a position for the imaging slab was chosen such that all vials and part of the volunteer's brain were contained in the 3D slab. With this setup several 3D data sets with different imaging parameters were acquired to demonstrate the influence of resolution and SNR, as well as the dependence of the texture parameters on different imaging parameters (eg, α,TR,TE). In a pilot study, texture parameters such as mean gradient show the same behavior in phantoms as in white matter for different patients, indicating that a normalization of texture parameters using test objects is possible ( However, texture normalization is necessary, but it is not possible to mimic all texture features by phantoms.[15]

Clinical application

The aim of this pilot study was to assess the possibility of quantitative description of texture directivity in trabecular bone with an attempt to quantitative description of trabecular bone structural anisotropy using texture analysis of 3D FLASH MRI. A series of 3D FLASH images, all of 256x256 pixels, with the voxel size of 0.4x0.4x0.4 mm3, were measured on a standard 1 .5-T scanner (Siemens Vision, Erlangen, Germany) using a small flex coil. The images in ( represent trabecular bone cross-sections in the sagittal and reconstructed transversal direction. For bone image texture analysis, circular regions of interest (ROI) were marked on corresponding bone cross-sections and effort has been made to maintain a large-size ROI for better statistical significance of texture parameters. The texture of the bone image shows apparent directivity, which reflects anisotropy of its physical structure according to the direction of gravity (Figure 8c). Quantitative analysis of this directivity is important to medical diagnosis, eg, in early detection of osteoporosis, as the directivity may vary according to the development, of the disease.