Cerebrovascular disease is a terrible medical and economic burden. Stroke, the most acute consequence of cerebrovascular disease, is a leading cause of death and disability in first world countries 1 . However, only about 10% of patients receive causal treatment, i.e. thrombolysis. On the other hand, developing chronic steno-occlusive disease, e.g. with atherosclerotic etiology or as an effect of Moya-Moya-Disease, is a diagnostic challenge, where only potentially harmful invasive diagnostic techniques are available to retrieve necessary diagnostic information. Consequently, diagnostic advances in cerebrovascular imaging are highly warranted. Neuroimaging plays a key role in this endeavor and magneticresonanceimaging (MRI) is a common neuroimaging method due to its inherent advantages: It is non-invasive, offers high spatial resolution and has no known long-lasting harmful effects on patients. Improved and new MRI neuroimaging methods have the potential a) to increase the validity of current diagnostic tools, b) to lead to new diagnostic tools and c) to replace current invasive and potentially risky diagnostic measures. The key question is, however, which diagnostic targets should be the aim of methodological advances. The international research community has developed a terminology to define diagnosis relevant imaging targets in acute stroke, so called “Treatment-Relevant Acute Imaging Targets” (TRAITS) 2 . These TRAITS serve as markers for inclusion or exclusion of patients into certain treatment protocols. This definition is beneficial, as it focuses imaging research on goals which are directly beneficial for patients. While TRAITS were predominantly developed for the use in acute stroke, a use in chronic forms of cerebrovascular disease such a steno-occlusive disease or Moya-Moya-disease is justified, as the imaging targets are similar. For the current thesis we defined three different TRAITS in three subprojects, where diagnostic improvements are warranted, and analyzed them in 7 publications:
7.4. RECONSTRUCTING SUSCEPTIBILITY DISTRIBUTIONS
7.4 Reconstructing Susceptibility Distributions
All strategies for estimating susceptibility distributions in soft tissue are based on the a priori estimation of fieldmaps, and thus on phase imaging. Single-echo GRE measurements are not useful for this aim, since fieldmaps derived of such data may contain deviations from the true field due to unconsidered phase offsets (see Section 6.1). Fieldmaps used for susceptibility recon- struction should at least rely on double-echo, or offset-corrected phase data. Furthermore it is evident, that the estimation of susceptibility requires knowledge of the magnetic field in all spatial dimensions. Ideally, an isotropic sampling of space should be attempted. This is best achieved using an isotropically sampled, full 3D GRE sequence with slab- or non-selective excitation. Sev- eral strategies for the reconstruction of tissue magnetic susceptibility were introduced during the last years. Two main categories can be defined, single and multiple orientation measurements. Furthermore, reconstruction can be performed directly or by using a minimisation approach. All methods presented below use the model of scalar susceptibility (Section 7.3.1), and will be shown in application examples using optimised parametrisation. Parameter optimisation will be discussed later on in Section 7.6. A comprehensive discussion of the current reconstruction techniques and applications for susceptibility imaging in MRI can be found in Reichenbach [ 2012 ].
[e.g., Ref. ( 18 )]. Of course, clinical fMRI—as all other applied neuroimaging techniques—requires clinical fMRI expertise and particularly pathophysiological expertise to be able to concep- tualize where to find what, depending on the pathologies of the given brain. One should be aware that full automatization is cur- rently not possible neither for a comparatively simple analysis of a chest X-ray nor for applied neuroimaging. In a clinical context, error estimations still need to be supported by the fMRI expert and cannot be done by an algorithm alone. As a consequence, the international community started early with offering dedicated clinical methodological courses (compare http://oegfmrt.org or http://ohbmbrainmappingblog.com/blog/archives/12-2016). Meanwhile, there are enough methodological studies that enable an experienced clinical fMRI expert to safely judge the possibili- ties and limitations for a valid functional report in a given patient with his/her specific pathologies and compliance situation. Of course, this also requires adequate consideration of the local hard- and software. Therefore and particularly when considering the various validation studies, neither for patients nor for doctors there is a need to raise “doubts about clinical fMRI studies” but instead good reason to “keep calm and scan on.” 4
such as lipid contamination and partial volume effects. Therefore, it is expected that a prospectively acquired low resolution data set will yield higher errors when reconstructed using the D-UNet. This must be evaluated in a more rigorous study where both low resolution and high resolution experimental SI data are acquired. Of course, the original resolution of the experimental SI plays a large role in the reconstruction process. While 24x24 and 32x32 matrices provide relatively accurate high resolution reconstructions, the 16x16 resolution does not perform as well. This suggests that there is a lower bound necessary to accurately upscale high resolution SI. This might be true for other super- resolution techniques ( 21 ), so a more thorough comparison between this deep learning method and other methods may aid in identifying this lower bound. Furthermore, results may be biased by the quantitative methods implemented to produce the LRSI before the super-resolution process is performed. This bias could be removed in the future by developing a deep learning based approach to metabolite quantitation ( 44 ). However, it may be worthwhile to explore the differences between common one dimensional spectral quantitation
The combined use of two unconventional NMR diﬀusometry techniques permits measurements of the self-diﬀusion coeﬃcient of ﬂu- ids conﬁned in porous media in the time range from 100 microseconds to seconds. The fringe ﬁeld stimulated echo technique (FFStE) exploits the strong steady gradient in the fringe ﬁeld of a superconducting magnet. Using a standard 9.4 T (400 MHz) wide-bore magnet, for example, the gradient is 22 T/m at 375 MHz proton resonance and reaches 60 T/m at 200 MHz. Extremely short diﬀusion times can be probed on this basis. The magnetization grid rotating frame imaging technique (MAGROFI) is based on gradients of the radio fre- quency (RF) ﬁeld. The RF gradients not necessarily need be constant since the data are acquired with spatial resolution along the RF gradient direction. MAGROFI is also well suited for unilateral NMR applications where all ﬁelds are intrinsically inhomogeneous. The RF gradients reached depend largely on the RF coil diameter and geometry. Using a conic shape, a value of at least 0.3 T/m can be reached which is suitable for long-time diﬀusion measurements. Both techniques do not require any special hardware and can be imple- mented on standard high RF power NMR spectrometers. As an application, the inﬂuence of the tortuosity increasing with the diﬀusion time is examined in a saturated porous silica glass.
3.2 Fast whole-heart imaging with 3D radial sampling 4
Fast volumetric imaging with isotropic resolution, such as that provided by the previously described 3D radial acquisition technique, is desirable in a number of clinical applications of MRI. One example, namely whole-heart coronary magneticresonance angiography (CMRA), is investigated in this section. This approach addresses a major shortcoming of coronary MRA, which is the complex and time-consuming examination, requiring extensive scout scanning prior to the actual image acquisition. These preparations are necessary to locate the coronary arteries, and to plan a particular measurement [stu99]. As an improvement, it is desirable to acquire an extended volume, e.g. the entire heart [web03], with high and isotropic resolution instead, making initial planning rather simple. Slices and arbitrary views can then be reconstructed retrospectively from the volume data set. In the present work, a 3D radial acquisition as described in the theory section was combined with a respiratory navigator-gated SSFP sequence to acquire a large volume covering the entire heart with high resolution during free breathing. Different views and slices containing e.g. the left anterior descending artery (LAD) and the right coronary artery (RCA) were reformatted retrospectively after completing the scans.
S. Haber-Pohlmeier,* C. Tötzke, E. Lehmann, N. Kardjilov, A. Pohlmeier, and S.E. Oswald
In situ investigations of the rhizosphere require high-resolution imaging tech- niques, which allow a look into the optically opaque soil compartment. We present the novel combination of magneticresonanceimaging (MRI) and neu- tron computed tomography (NCT) to achieve synergistic information such as water mobility in terms of three-dimensional (3D) relaxation time maps and total water content maps. Besides a stationary MRI scanner for relaxation time map- ping, we used a transportable MRI system on site in the NCT facility to capture rhizosphere properties before desiccation and after subsequent rewetting. First, we addressed two questions using water-filled test capillaries between 0.1 and 5 mm: which root diameters can still be detected by both methods, and to what extent are defined interfaces blurred by these imagingtechniques? Going to real root system architecture, we demonstrated the sensitivity of the transportable MRI device by co-registration with NCT and additional validation using X-ray computed tomography. Under saturated conditions, we observed for the rhizo- sphere in situ a zone with shorter T1 relaxation time across a distance of about 1 mm that was not caused by reduced water content, as proven by successive NCT measurements. We conclude that the effective pore size in the pore network had changed, induced by a gel phase. After rewetting, NCT images showed a dry zone persisting while the MRI intensity inside the root increased considerably, indicat- ing water uptake from the surrounding bulk soil through the still hydrophobic rhizosphere. Overall, combining NCT and MRI allows a more detailed analysis of the rhizosphere’s functioning.
(Figures 2, 3). Of note, this resolution and contrast was obtained without further injection of contrast agents, which is a clear advantage over CA-enhanced techniques for clinical translation. Still, contrast agents such as manganese or gadofluorine M were successfully applied to enhance resolution and contrast in pre- clinical PNS injury models ( 22 ). Thus, in a future study, it will be interesting to see whether isotropic high-resolution MRI protocols as established in this study results in further improved delineation of the nerves when injecting contrast agents or combined with further spin preparation techniques known to be sensitive to demyelination such as magnetization transfer contrast (MTC).
10.9. COMPARING THE METHODS
10.9 Comparing the Methods
When analysing phase data, one desires to employ a channel recombination method that generates robust results, but is also time-efficient. As long as the data fulfil the preconditions of e.g. the complex sum – meaning the absence of constant or spatially varying phase offsets in the channel data – one can use this method, since it is the quickest one. However, this is rarely the case. The SMW method is a quite robust choice that compensates for eventual phase offsets between the coil channels and offers a good compromise for everyday use. Further, SMW is utilised as part of the advanced methods for inhomogeneous phase offsets. When working with higher field strengths, it is advisable to pick a method such as DownCor or per-channel field estimation. Though both methods employ per-channel unwraps and estimate linear regression fits, DownCor is more time efficient, since it computes downsampled data. A downsampling factor of 1/4 of the source resolution implies massive speed-up for the unwrapping algorithm and thus for the entire offset estimation – the problem size is reduced to 1/64. Nevertheless, downsampling always leads to a loss in accuracy and to interpolation errors when resampling the offset data back to native resolution. When measurements contain phase contrast with high complexity, analysis at maximum accuracy is desired. Then, it is strongly advisable to choose the per-channel field estimation. Although the calculation of fieldmaps for each channel is highly time-consuming, one makes sure to keep the full resolution and all information contained in the measured dataset. In addition to that, novel techniques such as the URSULA algorithm presented in this thesis (see Chpt. 5) allow for significant speed-up and make the per-channel field estimation feasible within hours or minutes, even for high-resolution data.
Another post-processing strategy frequently applied in functional connectivity analysis is low pass or band pass temporal filtering of the rsfMRI signal. It removes high frequency Gaussian noise and in turn emphasizes low frequency fluctuations around 0.1 Hz. We found that the frequency bands contained in our respiratory and cardiac regressors are relatively broad. Due to variations and high undersampling of the respiratory rate, especially the position of the respiratory band varies from subject to subject and may interfere with the low frequency fluctuations of interest. Hence, it is clear that plain low pass or band pass filtering cannot remove these physiological fluctuations, a conclusion earlier reached also for fMRI of the macaque (Teichert et al 2009). Some approaches using band stop filters selective to frequencies of physiological relevance have been successfully applied (Biswal et al 1996; Buonocore and Maddock 1997; Deckers et al 2006). These techniques are computationally less demanding compared to linear regression, but require comparable physiological data (and may still interfere with the low frequency spectrum).
Publikation 3: Tahir Durmus, Carsten Stephan, Maria Grigoryev, Gerd Diederichs, Musaab Saleh, Torsten Slowinski, Andreas Maxeiner, Anke Thomas, Thomas Fischer; Detection of prostate cancer by real-time MR/Ultrasound fusion guided biopsy: 3T MRI and state of the art sonography techniques; Rofo. 2013 May;185(5):428-33.: Dateninterpretation. Verfassen einzelner Passagen und kritisches Überarbeiten des Manuskripts. Finale Zustimmung der zu publizierenden Version des Artikels
With high-field-strength magnets, the problems corresponding to susceptibility effects will become very severe, particularly at 17.6 T. For this reason, a spin-echo sequence, less sensitive to susceptibility effects than echoplanar imagingtechniques, is preferred (Basser 1994; Bernstein 2004). However, the spin-echo sequence is highly sensitive to motion. Motion could arise from two sources: patient or sample bulk motion and vibrations of the sample apparatus induced by the strong diffusion-weighting gradient pulses. In our experiments, there was no patient motion and it has been taken care that gradient-induced vibration was avoided through mechanical stiffness of the measurement device. The spin-echo sequence is also characterized by very long data acquisition time, but this point is decisive in the first instance only for clinical applications.
patients and controls are in line with a study using FS 4.3 and a mixed antidepressant treatment paradigm ( Phillips et al., 2015 ). Likewise, the authors did not find baseline hippocampal differences between depressed patients and healthy con- trols. Importantly, this was a more severe treatment-resistant sample undergoing antidepressant therapy with an average scan interval of 331 days in remitters and 420 in nonremitters. Moreover, Philipps et al. found that patients who did not remit over the course of the treatment period exhibited larger base- line volumes. Only patients who remitted exhibited volume in- creases according to treatment (n = 26, 1.5T, 12 months) ( Phillips et al., 2015 ). Larger baseline volumes in patients who do not remit in are in concert with our results; however, we did not replicate their positive remission × time interaction in the whole hippocampus in our subsample. Increased baseline volumes, however, speak against a meta-analytic finding, indicating that smaller baseline volumes are associated with lower response/ remission rates ( Colle et al., 2018 ). At this stage, it is only pos- sible to speculate about reasons for these findings. Studies incorporating disease-inherent heterogeneity with methodo- logically standardized measurement techniques at comparable time points of the disease phases are needed to better compare hippocampal volume studies.
A major technical challenge in coil array design is the mutual decoupling between individual coil elements. Conventional decoupling techniques use either geometrical overlap , with the drawback of reduced overall FOV and higher g-factors for parallel imaging due to the overlapping sensitivity profiles. Another decoupling strategy includes LC-networks between nonoverlapping coils [63,65], with the disadvantage of frequency and load-dependent decoupling efficiency. Some authors proposed strategies to decouple physically separated coils by magnetic flux sharing to combine the advantages of overlap and LC-network decoupling. Avdievich and Hetherington  used a pair of overlapping annex loops with opposite winding orientation connected in series with two neighboring surface coil elements. Constantinides and Angeli  placed closed copper loops proximal to the array, partially overlapping with the mutually interacting surface coils, and thereby eliminating the magnetic coupling. Low-impedance preamplifiers are widely used for inter-element decoupling in receiver arrays . In transmit arrays, the mutual coupling can be reduced with the current source RF amplifier method [61,62], although it is currently not available for most MRI systems.
nation, but the distribution of the catalyst system on the monolith is challenging to reveal. For surface-sensitive analytic techniques, such as scanning electron microscopy (SEM) or transmission elec- tron microscopy (TEM), the monoliths need to be divided into small samples, which might inuence the catalyst system. To this end, less invasive techniques are therefore preferred. X-ray absorption techniques, such as computed tomography (CT), are sensitive to heavy elements and therefore lack in contrast between IL in the pores and the solid SiC matrix. For high resolution, micro- tomography (micro-CT) is favorable but also requires invasive cutting of the monolith. An alternative, non-invasive technique with lower resolution focusing on the liquid catalyst system is magneticresonanceimaging (MRI) being one of the modes of nuclear magneticresonance (NMR). MRI applies radiofrequency radiation and provides therefore access to optically opaque materials while unique contrast parameters such as NMR relaxation, can be exploited to increase the information content. Accordingly, MRI has been used for the examination of a wide range of both, so and hard materials such as polymers, food, plants and wood 18–22 as well
12 OUTLOOK 12.3 AI-Based Motion Measurements and Correction
of current AI-based motion correction methods is how it can be guaranteed that they do not interpret pathologies as motion artefacts and mistakenly remove them. For obvious reasons, the resulting images in such a case would have a significantly reduced diagnostic value with potentially catastrophic consequences for the patient. A possible solution for these two problems could be SNAV-based motion estimates that serve as a guidance or "regularisation" for the AI-based motion correction techniques, limiting the scope of permitted corrections. For instance, if the SNAVs measured a motion pattern as the one in the uppermost graph in Figure 9.5, i. e. virtually no motion, the freedom of the AI to correct motion artefacts in the data would be restricted to a minimum, thereby preventing the mistaken removal of a pathology. If, in contrast, the measurements showed heavy motion like the one in the lowermost graph of Figure 9.5, the AI would be granted a greater freedom for correcting the corresponding artefacts. Importantly, the AI could be forced to find a motion correction approach that is consistent with qualitative details of the SNAV measurement, e. g. the motion type (rotation, translation, periodicity etc.) or the direction (through-plane, in-plane etc.). This would again assure that the important and real features of an image are left untouched and only motion-induced changes are undone.
Despite the numerous advantages associated with MRI, the major limitation is the exclusion of ferromagnetic metallic objects such as implants, which may be attracted by the strong magnetic field, making them a serious danger when placed within a strong magnet. Also, as the signal is based on hydrogen nuclei in most cases, arid structures such as bones cannot be imaged well. In terms of patient comfort, as the MR scanner usually has a rather small bore, measurement of claustrophobic patients is more complicated. However, in sometimes open bore MRI scanners are used, which offer more freedom and overcome this problem. Danger associated with radio frequency (RF) burns induced by secondary RF fields (cf. section 2.2.3), are usually minimised by safety constrains of the hardware. Compared to most other imagingtechniques, MRI is relatively expensive, mainly due to the requirement of liquid helium, which cools the superconducting magnet, and acquisition times are usually longer than in other imagingtechniques, requiring the patient to hold still for several minutes. To ensure the safety of patients during MR imaging, monitoring of the specific absorption rate (SAR) was introduced, which means measuring the amount of energy deposited in tissue (units: W/kg).
Chapter 3. Functional MagneticResonanceImaging (fMRI)
so-called fat suppression techniques (cf. section 8.2).
3.2.5. N /2 Ghost
A specific EPI artefact is the so called N /2 ghost caused by a misregistration between odd and even lines acquired. The N /2 ghost is a rather faint duplicate of the imaged object shifted by half the field-of-view. The intensity of the N /2 ghost is usually modulated by a sinosoidal signal profile in readout direction which reflects the phase difference corresponding to the shift between the k-space lines. Various methods have been proposed to correct for these ghosts, some of which do not even require a separate calibration scan. The most prevalent calibration method relies on the acquisition of two non-phase-encoded projections using a very short echo train immediately following excitation (and slice rewinding). Assuming, without loss of generality, the first readout is positive, the second, negative readout is usually followed by an additional third positive readout resulting in three projections with extrememly short echo time and thus very little phase evolution (almost no signal drop-outs, etc.). The first and the third projection, i.e. the Fourier transformed odd numbered signals, are averaged and the complex phase is compared to the complex phase of the second projection.The systematic phase difference can then be corrected for in the following imaging scans. Unless stated otherwise, for all results shown in this work a linear approximation of the phase (constant offset and slope) is used for phase correction. This leads to reasonable N/2 ghost reduction, at least at moderate field strengths of 3 Tesla. Alternative methods include non-linear phase correction and a pixel-by-pixel phase correction which can be computed from standard phase correction scans as well. A detailed discussion of these fundametal phase correction methods can be found in Ref. .
of animals required. Depicting hippocampal volume using magneticresonanceimaging (MRI) is a frequently used method in this context. A Medline search in February 2011 using the Medical subject heading (MeSH) search terms “hippocampus” combined with “MRI” and “shape” or “volume” extracted 1864 publications. Modern imagingtechniques in humans enable reliable measurement of hippocampal volume using native MRI contrast by defining hippocampal boarders using gray and white matter contrast differences (Konrad et al., 2009). Over the last decade high resolution 3D-images of mice, providing good structural insight, have been made possible by the development of adequate data acquisition protocols for small animal MRI (Natt et al., 2002). Despite these technical advances, few studies use native contrast for hippocampal volume delineation, which define hippocampal boarders manually (Maheswaran et al., 2009; Radyushkin et al., 2010). Low grey-white matter contrast in native MR images of mice makes it difficult to differentiate regional boarders. To enhance image contrast, brains are fixed and scanned using high field scanners and long MRI scans, a method called magneticresonance microscopy (MRM) (Badea et al., 2007). However this method does not utilize the great advantage of MRI, namely the possibility to repeatedly assess brain morphology in vivo. To overcome low regional intensity differences in vivo, contrast agents can be applied (Mendonça-Dias et al., 1983). The manganese ion Mn 2+ turns out to exhibit very promising
Several ultra-high field specific techniques for acceleration of 2D- and 3D-EPI have been pro- posed and implemented on a 3 Tesla and a 9.4 Tesla human MRI scanner. At both field strengths the custom sequences have achieved higher acceleration factors or less geometric distortions or highly reduced specific absorption rates compared to what the vendor provided sequence for BOLD functional imaging is capable of. In particular the proposed fat suppression method by means of a single on-resonant rectangular pulse opened up new opportunities for high reso- lution whole-brain functional MRI at ultra-high fields such as 7 or 9.4 Tesla. By omitting the conventional fat-saturation module prior to each excitation the specific gain with the proposed method is threefold: first, a “dead time” of more than 10 milliseconds per shot with the con- ventional method is eliminated; second, SAR is reduced from unacceptable levels preventing the scanner to run a BOLD fMRI protocol under realistic conditions to exceptionally low values without losing fat-suppression efficiency; third, additional water signal suppression due to an unwanted magnetisation transfer effect with the conventional method, that appears to be quite large when utilised in 3D-EPI sequences, is completely avoided. Overall, an exemplary increase of sensitivity (tSNR/ p acquisition time) has been achieved. Utilising additional physiological information (respiration, heart beat) is expected to even more improve 3D-EPI raw data and/or functional data.