Tim Schakel

Generally, there are two approaches to reduce the distortions in DW-EPI during acquisition. The first is to increase the phase encoding bandwidth, this can be achieved by reducing the number of acquired lines in the phase encoding direction. This easiest way to do this is sacrificing spatial resolution; although generally this is undesirable due to the already limited spatial resolution in single shot EPI. However, when using parallel imaging, the image can be reconstructed in the full resolution while acquiring less lines. This speeds up the acquisition and increases the phase encoding bandwidth. This does come at the cost of increased noise in the images. Additionally, the amount of acceleration is limited by the amount (and configuration) of receiver coils in the imaging setup. In the current coil setup, where patients are scanned in their immobilization mask, the amount of parallel imaging is already maximized. The development of coils with a large number of channels [126], which are compatible with the immobilization mask, can further improve image quality. Not only to allow for higher parallel imaging factors [127], and consequently reduced image distortions in DW-EPI, but also to improve SNR.

The second approach is to decrease the underlying magnetic field inhomogeneity. This can be done through shimming. In active shimming, specialized shimming gradient coils are used to generate a corrective magnetic field to counteract the subject induced magnetic field inhomogeneity [128, 129]. Modern 3.0 T MR systems have shimming gradients that can produce second order corrective field, while 1.5 T MR systems can typically only produce first order fields. In order to calculate the appropriate corrective shim field, some estimate of the magnetic field inhomogeneity is required. Once this field inhomogeneity is known, an optimization algorithm can be used to derive the shim fields [130]. This can be done for the entire field of view, or optimized on a specific region of the anatomy [131]. In the latter case, the magnetic field inhomogeneity of areas not included in the shim optimization algorithm could worsen. Several methods have been proposed to further improve optimization algorithms, using additional constraints in the algorithm [132, 133] or by utilizing information on the location of water and fat in the image [134, 135]. First, shimming on water only will improve the shim by reducing the off-resonance effect of fat. Second, information of water and fat resonance frequencies can be used to optimize shims for spectral fat suppression, since the magnetic field homogeneity also influences the effectiveness of spectral fat suppression methods. However, due to its proximity to the air tissue interface, the magnetic field inhomogeneities around a head and neck tumor are generally more complex than can be corrected for using second order shim gradients. An alternative approach would be to use passive shimming, e.g. by placing a susceptibility-matched material around the patient. While this alleviates some of the magnetic field inhomogeneities around the complex chin-neck-shoulder shapes, the issues with the internal air-tissue transitions remain.

Turbo spin echo

Alternatively, diffusion weighted MRI can also be acquired using a single shot turbo spin echo (TSE) readout [28–30, 32, 35]. In single shot TSE, RF refocusing pulses are inserted between every readout line. These RF refocusing pulses will reverse phase accrual from off-resonances (such as magnetic field inhomogeneities or eddy currents), removing the phase encoding errors that are observer in EPI sequences. The geometric accuracy of DW-TSE sequences is comparable to that of standard anatomical sequences.

The DW-SPLICE sequence was implemented as alternative for DW-EPI in chapter 3. Using the DW-SPLICE sequence diffusion weighted images with excellent geometric accuracy and good overall image quality can be acquired. The geometric distortions that DW-EPI produces in the phase encoding direction are not present in TSE based sequences such as DW-SPLICE. Additionally, the bandwidth in the readout direction was large enough such that any distortions in this direction, i.e. due to off-resonance effects, are on the sub millimeter level. This means these images are suitable for target volume delineation and use in radiotherapy treatment planning. An example is also given in figure 7.1, this is the same patient as in the introduction of this thesis (figure 1.2 and 1.3).

The main disadvantage of DW-SPLICE, and TSE methods in general, is the increased acquisition time. TSE methods are inherently slower due to the inclusion of the RF refocusing pulses. In the results presented in chapter 3, it takes 16.4 s to acquire a single volume of diffusion data with DW-SPLICE, while this is 3.3 s with DW-EPI. This limits the amount of data (i.e. b-values or signal averages) that can be acquired within a clinically acceptable time. In practice this means less b-values will be acquired, the signal averages are required in order to have the diffusion weighted images have sufficient SNR. The reduction of the number of b-values limits the diffusion quantification possibilities. Simple models, such as a mono exponential fit to estimate the ADC value can be performed using 2 or 3 b-values. However, more sophisticated methods such as the bi exponential intra-voxel incoherent motion model require more b-values [109]. As DW-SPLICE is a single shot method, it has a limited spatial resolution. Increasing the resolution would increase the number of lines acquired per shot and increase the shot length. This would mean a loss of echo signal along the shot due to the T2 relaxation of the transverse magnetization. This decrease in signal over the echoes also acts as a k-space filter, resulting in blurring in the reconstructed images that increases with shot length.

Figure 7.1: Diffusion weighted MRI, acquired with DW-SPLICE, with different diffusion weightings and a calculated ADC map. The primary tumor exhibits diffusion restriction; it shows high signal intensity on the diffusion weighted images and low values in the corresponding ADC map. Top row: b = 0 s/mm2 and b = 200 s/mm2, bottom row: b = 800 s/mm2 and ADC map.

In order to further improve the DW-SPLICE image and decrease the amount of blurring, multi shot methods are an option. Since less lines are acquired per shot, the shot length decreases and with it the amount of blurring. This could also facilitate the acquisition of images at a higher spatial resolution. However, multi shot acquisitions of diffusion weighted data are challenging. This is due to the same problems that hindered TSE acquisitions, motion-induced phase errors. These phase errors will change with every shot making it difficult to align the shots. Some form of phase navigation is required in the multi shot methods. This is typically done using 1D navigators, however, in the case of DW-MRI, these might not suffice because the phase errors have a more complex pattern. To estimate the phase error, a 2D phase navigator would have to be acquired. An alternative approach would be to use a non-cartesian k-space trajectory: PROPELLER (periodically rotated overlapping parallel lines with enhanced reconstruction) [136]. With PROPELLER, a set of adjacent lines (blade) is acquired in the center of k-space with every shot. This blade is then rotated between the shots, such that a common circle of k-space is acquired by all the blades, but every blade also has unique high spatial frequency information from the outer k-space. PROPELLER is a self-navigating method, the common circle of k-space data per blade can be used as navigator data. Combining DW-SPLICE with a PROPELLER acquisition scheme could provide a gain spatial resolution but also decrease the blurring [34, 81].

Another much needed improvement is acquisition speed. Generally, increasing the acquisition speed, without sacrificing spatial resolution, is done by applying parallel imaging. However, with the current coil setup, the parallel imaging is already maximized. The development of a dedicated receive coil setup with more elements, which is also compatible with the RT immobilization mask, could enable higher parallel imaging factor to speed up the acquisition. Besides the traditional acceleration method, compressed sensing (CS) techniques [137, 138] have been recently introduced in MRI. Following a semi-random undersampling of k-space an iterative reconstruction is performed in order to generate the final image. CS exploits sparsity, the information can be represented using a limited number of non-zero coefficients. The sparsity can be in the image directly, or obtained through some sparsifying transform. Diffusion data seems very well suited for CS, the images themselves are relatively sparse. But a series of corresponding diffusion weighted images is also sparse, the amount of information that changes between the differently diffusion weighted images is limited [139–141]. CS has been shown to enable very high acceleration factors, up to factors of 10 or 20. The disadvantage of CS is the reconstruction time, while the acquisition could be greatly sped up, the iterative reconstruction can take hours.

Diffusion preparation

The EPI and TSE methods discussed so far are slice selective methods, they apply a slice selective excitation after which the signal is spatially encoded for this slice. 3D acquisitions have an intrinsic higher SNR, making them very promising for diffusion acquisition. To facilitate 3D diffusion acquisitions, driven equilibrium sequences have been proposed [142, 143]. Following diffusion preparation a 90 degree RF reset pulse is give, tipping all diffusion weighted magnetization to the longitudinal axis. This magnetization can now be excited to perform a 3D readout. This method has been shown with TSE readouts but also with gradient echo readouts (bSSFP, TFE) [144–146]. An added benefit of using 3D over 2D acquisitions is their improved compatibility with CS acceleration, more easily allowing undersampling strategies.

Fat suppression

The acquisition of DW-MRI is generally accompanied by fat suppression. Unsuppressed fat will appear bright on a diffusion weighted image and has a low ADC value, characteristics similar to tumors. The presence of fat can obscure tumors, degrade image quality and make images harder to interpret and process. Furthermore, unsuppressed fat can shift in the images as a function of the pixel bandwidth, due to the chemical shift between water and fat. In DW-EPI sequences this effect is especially pronounced because of its low pixel bandwidth. Fat shifts up to half the field of view are not uncommon here. So even if the tumor is not in close proximity to fat, unsuppressed fat from other parts of the image can shift on top of it.

In chapter 5, the DW-SPLICE sequence was combined with water fat separation in order to achieve more robust fat suppression in the head and neck region. The DW-SPLICE sequence was modified to acquire additional data at different echo times, generating a data set that was suitable for water fat separation. Water fat separation was performed on the b0 images and then applied to the b800 images to remove the fat. The water fat separation could not be performed in the diffusion weighted images directly due to the complications of the previously discussed phase errors. The combination of DW-SPLICE and water-fat separation showed promising results in improving the fat suppression in the head and neck region. Further improvements could also enable direct water fat separation on the diffusion weighted images. Recently, several authors have reported on hybrid techniques that process complex and magnitude-only data sequentially [114, 120].

7.2 Diffusion weighted MRI in radiotherapy

Target volume delineation

The sensitivity of DW-MRI to tissue characteristics such as cell density and tissue organization make it a very promising technique to use in target volume delineation. And the excellent geometrical accuracy of the DW-SPLICE sequence allows for the acquisition of diffusion weighted images that can be used in radiotherapy treatment planning. The information in a diffusion weighted scan is more sparse than in anatomical MR scans, implying that diffusion weighted scans are more straightforward to interpret. This is supported by the high conformity index of delineations on DW-MRI as reported in chapter 4. Three observers had a conformity index of 0.73, indicating good agreement. The inclusion of DW-MRI could help in decreasing delineation variation between different observers.

The sparse nature of the diffusion weighted images also makes them suitable for automatic processing. In chapter 4 a simple method to segment a region in the b800 image using an image intensity threshold of 50% was applied. Delineators then manually adapted this initial segmentation based on information from the ADC map and the rest of the diffusion weighted images. Further development of the automatic segmentation methods should utilize all of the images in the diffusion series. Machine learning methods, such as support vector machines [147] and neural networks [148, 149] seem very well suited for this. The development and training of such methods could benefit greatly from matched gold standard pathology data that can act as a ground truth [41, 42].

From previous validation studies it was found that automatically segmented PET provides a tumor volume closest to the volume found on histology with a high coverage and the lowest overestimation. Chapter 4 compares delineations on DW-MRI to automatic segmentations of PET. The volumes found on DW-MRI (10.8, 10.5 and 9.0 cm3 respectively) were larger than those on PET (8.0 cm3). The difference was significant for two out of the three observers. There was a substantial overlap between DW-MRI and PET with DSCs of 0.71, 0.69 and 0.72 respectively, indicating that, for a large part, both techniques identify the same target. The DW-MRI and PET volumes were considerably smaller than the clinically used GTV (15.6 cm3) for these patients. However, it is also known, from validation and interobserver studies, that the true GTV tends to be overestimated. The GTVs as segmented on PET still require a CTV margin in order to encompass all tumor tissue [85]. Likewise, the GTVs from DW-MRI will also need an appropriate CTV margin.

This is also indicated by the initial results of an imaging validation study with DW-MRI as described in chapter 6. This imaging validation study continues the one initiated by Caldas-Magalhaes et al. [41] but includes MRI at 3.0 T with an updated imaging protocol, also containing DW-SPLICE. The results of this first patient showed very good correspondence between DW-MRI and pathology, with volumes of 11.7 and 12.3 cm3 respectively. Besides the delineations showing similar volumes, the delineations also had a large spatial overlap as evidenced by a dice similarity coefficient of 0.82. Moreover, the sensitivity and positive predictive value of DW-MRI were high with 0.80 and 0.84 respectively. These first results for DW-MRI are encouraging. Currently, further patient inclusion for this study is ongoing and its results can provide valuable insights in the interpretation of imaging [121].

The addition of DW-SPLICE to the current practice of target volume delineation can help reduce variation among observers and increase target volume delineation accuracy. (Semi-)automatic segmentations can provide a first estimate for the tumor outline. This first estimate can be combined with the segmentation of PET as the region identified by both techniques will be of great importance for treatment.

Treatment response monitoring

In addition to target volume delineation DW-MRI can be used evaluate the response to treatment. Successful treatment should result in a loss of cell membrane integrity and a loss of cellularity in general. These changes can be detected as an increase in the ADC values of the tumor. Several studies have investigated the use of DW-MRI for response monitoring and prediction and have showed promising results [112, 150, 151]. Most studies evaluated changes in the mean ADC of the whole tumor at different time points during treatment. But it is also possible to evaluate the changes in ADC on a per-voxel basis [48]. This does require diffusion weighted images with good geometric accuracy to enable image registration between the different time points. DW-SPLICE, acquired in an immobilization mask, is very well suited for this. Currently, such a study is ongoing at our department.

7.3 Future perspectives

Advances in MRI acquisition

The field of MRI is still very much in development. Advances in hardware and software enable new possibilities and applications. There is an increasing trend in further acceleration of MR acquisitions. Undersampled datasets are reconstructed to full resolution images using advanced reconstruction algorithms. One of the most prominent methods is CS [137, 138]. The additional functional dimension in DW-MRI datasets can be exploited by a CS reconstruction to further accelerate data acquisition. As the MR reconstructions tend to become more complex and computationally demanding, the resources on current clinical MR systems are often insufficient. Therefore, external reconstruction resources, that automatically process acquired data pushed from the MR system, will be very beneficial, not only in developing and prototyping new methods but also to help clinical introduction of promising methods. Non-cartesian acquisition strategies such as PROPELLER can enable higher spatial resolutions. Moreover, these strategies also synergies very well with CS reconstruction methods.

Imaging patients in radiotherapy treatment position, in their customized immobilization masks, limits the type of coils that can be used. Currently, flexible surface coils with two large elements are used. The development of a dedicated coil, compatible with the immobilization mask and with an increased number of receive elements would allow to further exploit parallel imaging strategies and increase SNR.

The imaging for radiotherapy treatment planning in head and neck cancer is multimodal. Taking this into account, another interesting hardware development is the introduction of hybrid PET-MR systems, which allow for the simultaneous acquisition of PET and MR data [152–154]. PET-MR can provide a combination of anatomical, functional and metabolic information in a single exam without misalignment or a time delay between the different acquisitions. The relatively long acquisition time of MR images could be beneficial for the image quality of the PET due to this. Additionally, multimodal reconstructions have been proposed to improve the images of both modalities [155]. An important requirement for PET-MR are electron density maps, for use in correcting the PET data but also in dose calculations in treatment planning. These would normally be derived from CT, but now have to be derived from MR data.

Towards daily DW-MRI

The recent introduction of MR-Linac systems enables MR imaging at the time of patient treatment [156]. This provides unique opportunities to perform daily DW-MRI scans. Monitoring changes in the diffusion characteristics of the tumor could help in treatment response assessment and indicate areas in which treatment adaptation would be required. This does require the translation of the results in this thesis from 3.0 T to 1.5 T. Getting sufficient SNR in DW-SPLICE acquisitions on the MRL would most likely be the biggest challenge. The TSE methods benefit from the increased SNR at higher field strengths. Additionally, the MRL has reduced gradient performance when compared with the 3.0 T MR system. This means achieving a certain b-value will take more time and will consequently increase the echo time and lower the SNR. System characterization in terms of eddy currents and gradient performance will allow for further optimization of DW-MRI on the MR-Linac. Promising results in terms of diffusion quantification have already been acquired. Initial ADC measurements using a standardized phantom show excellent correspondence with the calibrated values without image distortions [157].