Background
Radiotherapy for upper abdominal cancer patients, such as pancreatic, esophageal and gastric cancer, is challenging because of poor contrast between these tumors and other soft tissues on planning CT scans and because of respiratory-induced tumor and organ motion. It was shown that there is a large variation between observers when delineating upper abdominal tumors on computed tomography (CT) [
1]. This variation appears to decrease when adding MRI images, which offer better contrast between tumor and the surrounding tissues [
1‐
3]. Respiratory-induced tumor motion, during irradiation is another major challenge for abdominal cancer. Using an internal target volume (ITV), derived from a four-dimensional CT (4DCT), is one way to take this into account. However, for pancreatic cancer it was shown that the initial 4DCT is not representative for the respiratory-induced motion during actual treatment, whilst for esophageal cancer this is the case [
4,
5]. Furthermore, to limit X-ray dose burden, the 4DCT acquisition is typically done with a limited number of acquisitions, compromising image quality even further.
4DMRI modalities are promising to overcome the challenges mentioned above, as 4DMRI has superior soft tissue contrast, is flexible, allowing for various image acquisitions, and omits additional imaging radiation dose. A T2-weighted contrast is desirable for tumor and organ at risk visibility [
6,
7]. Finally, contrary to CT, MRI offers means for measuring the respiratory signal using internal surrogates, which is shown to be more accurate for amplitude binning strategies compared to the use of external surrogates [
8].
One of the challenges of 4DMRI (and 4DCT) is that irregular breathing is known to deteriorate image quality and introduce image artifacts [
9‐
15]. Various strategies of handling irregular breathing have been demonstrated for 4DMRI: e.g., discarding of images associated with an amplitude that falls outside a defined range of amplitudes or applying triggering in order to only obtain images within a certain respiratory amplitude range [
11,
16,
17]. However, so far no studies were performed comparing the 4DMRI quality before and after outlier handling. In the presence of irregular breathing, it is vital to balance the exclusion of outliers: including too many data allows outliers to degrade the image quality, whilst excluding too many data potentially leads to the 4DMRI only being representative for a limited percentage of the breathing cycle. Therefore, an important aspect is how the motion of organs and tumors in the reconstructed 4DMRI after outlier handling should be interpreted in radiotherapy.
Ultimately, for radiotherapy treatment planning, a 4DMRI is preferably artifact-free, precise (i.e., the image should accurately depict the anatomy) and give an accurate measure of the respiratory-induced motion during imaging. 4DMRI reconstruction quality should be evaluated on exactly these features. However, most reports on 4DMRI reconstruction quality scored the 4DMRI visually [
11]. Some quantitative work has been done, such as fitting a smooth curve to the organ border irregularities and assessing the amount of missing data after retrospective image sorting [
13,
18]. Other studies validated 4DMRI reconstruction accuracy with phantom measurements where the reconstructed motions were compared to the physical motion of the phantom [
14,
16,
19]. For radiotherapy purposes, all of these quantitative quality measures (i.e., artifacts, precision and validation of reconstruction accuracy) should be addressed.
In this study, we present a 4DMRI acquisition and reconstruction optimized for radiotherapy treatment planning in the abdominal region. We developed a novel outlier rejection strategy and combined it with amplitude binning for robust 4DMRI reconstruction in the presence of irregular breathing. We validated the 4DMRI reconstruction with a dynamic phantom study. Furthermore, by defining a set of 4DMRI quality parameters the precision of the diaphragm position, accuracy of motion amplitude, and the occurrences of image artifacts were quantitatively assessed making it possible to compare 4DMRIs in an objective way. Finally, we compared our outlier rejection strategy and amplitude binning with three other outlier rejection and binning strategies.
Discussion
We presented a 4DMRI acquisition and reconstruction approach, with robust and precise 4DMRI reconstruction (i.e., fewer artifacts) in the presence of irregular breathing compared to other binning strategies. The acquisition time of 6 min is suitable for clinical routine and provides information about patient-specific breathing motion over extended duration. Validation and quality assessment of the produced 4DMRIs was provided in terms of diaphragm position precision and artifact occurrences. This is the first 4DMRI study providing quantitative assessment of reconstruction precision, image artifacts, and motion validation comparing different binning and outlier rejection strategies, supplemented with a phantom validation study.
Commercially available MRI sequences were used, followed by an in-house developed post-processing technique that sorted scanner-reconstructed images. Thus, adaptations at the scanner such as customized sequences and patched software installations were not required. As a result, the image quality was up to clinical standard and can be used in clinical routines. The resulting images provided information on patient specific respiratory-induced motion, which can be incorporated in radiotherapy, e.g. by using a personalized internal target volume definition or algorithms for mid-ventilation techniques [
22,
23]. The superior soft tissue contrast, optimized for imaging organs in the abdomen, and high quality images (few artifacts, precise imaging) may reduce delineation uncertainties compared to CT scans [
1,
2], potentially reducing planning target volumes and thereby healthy tissue irradiation and corresponding toxicities.
Twelve healthy volunteers and two patients with cancer in the upper abdomen were included in this study. Although the volunteer cohort might not entirely represent the patient population’s respiratory characteristics and its variation, the volunteer cohort showed a large variety of breathing irregularities (see Additional file
1: Figure S4).
The presented 4DMRI acquisition consists of coronal images, since this depicts the diaphragm dome optimally, allowing for quality assessment of the resulting 4DMRI. The use of the well-established T2-weighted single shot turbo spin echo imaging benefits from high in-plane resolution. Furthermore, each slice was acquired in a short time span of 551 ms per slice providing a sharp snapshot of the anatomy during free breathing. In this study the diaphragm position in the coronal slices was used for median image selection. This could have been done on diaphragm position extracted from the 1D navigator as well, enabling the 4DMRI reconstruction as described in this manuscript to be done for transversal and sagittal slices as well.
The presented 4DMRI acquisition and reconstruction is a result of the requirements of radiotherapy and has advantages and disadvantages compared to other reported 4DMRI methods. The choice for adopting retrospective binning is analogous to conventional 4DCT, which facilitates implementation in current clinical practice. In contrast to 4DCT, 4DMRIs can be reconstructed in many more ways other than using an external surrogate for registering the respiratory signal. For instance, retrospective binning in k-space provides an elegant means to reconstruct 3D volumes by combining acquisitions of k-lines from different respiratory cycles, benefiting from fast acquisition and high signal-to-noise images [
8]. However, these k-space sorted 4DMRI techniques are limited to steady-state sequences, limiting the contrast to T1-weighted or T2/T1-weighted in most cases [
24,
25]. To have more T2-like weighted (T2W) contrast requires more complex steady-state sequences not commonly used for abdominal imaging [
6]. Alternatively, the deformation field from a k-space sorted T1-weighted 4DMRI can be applied to a 3D T2W contrast image [
26]. An improvement in 2D reconstruction could be achieved by acquiring 2D images perpendicular to the coronal planes and apply a super-resolution reconstruction method in order to reconstruct high-resolution 4DMRI with T2W contrast [
27]. Furthermore, evaluating the precision of the resulting 4DMRIs cannot be done on for instance diaphragm shape artifacts (i.e. S), relying on a 2D acquisition technique rather than a 3D acquisition. The outlier rejection strategy described in this work may also be applied for k-space binning, potentially improving the image quality further. However, as most k-space binning approaches rely on central k-space magnitude instead of absolute diaphragm position, the technique might need slight adapting for best results.
A 1D navigator has been shown to be a better surrogate for abdominal organ motion than an external surrogate [
8]. For amplitude binning, correspondence of surrogate displacement with organ displacement is more vital than for phase binning. For this reason, the use of a navigator instead of external surrogates [
14,
16,
28] to determine the respiratory signal was preferred for amplitude binning based 4DMRI reconstruction. Self-navigation, e.g. using the diaphragm position on the acquired 2D images for breathing registration is a promising option though outside of the scope of this study [
29,
30].
The presented scanning sequence acquired 60 dynamics (i.e., repetitions of the scanned volume) in 6 min. When sorting the images, a bin-slice combination may contain multiple images; this we exploited in quantifying the reconstruction precision. However, by selecting the image with the median diaphragm position as the image for 4DMRI reconstruction, an underestimation of motion amplitude was introduced. This underestimation was quantified by phantom measurements and was conform the calculated expected underestimation. For 4DMRI, similar errors in measured amplitude are reported, and in a phantom experiment a similar relative motion amplitude underestimation of 5% was found for 4DCT [
15,
25].
With the presented 4DMRI reconstruction method in clinical practice, motion amplitude underestimation will occur and can be expected to be in the range of 0.6–5.4% of the motion amplitude if one uses 10 respiratory bins and assumes that the peak and valley shape are between very sharp (i.e., the peak of a cos
6 waveform) or rather flat (e.g., the plateau of a cos
6 waveform), see Additional file
1: Figure S3. Using more and smaller respiratory bins for 4DMRI reconstruction, which may come at a price of lower RC or longer scan times, reduces the motion amplitude underestimation.
Our novel amplitude binning strategy adopted outlier rejection, preventing that outlier positions define the range over which amplitude binning was performed, which would result in poor RC or IBV. We chose to discard 5% of the data to mitigate the largest effect of irregular breathing, resulting in a mild underestimation of motion, directly related to the fraction of time the 4DMRI is in fact correctly depicting the organ motion trajectory. On average, discarding 5% of the data resulted in a 28% smaller IR, which may lead to a smaller patient-specific internal target volume, potentially sparing healthy tissue. The ITV is designed to cover the entire tumor trajectory, though this assumes that the 4D imaging provides the correct tumor motion. Irregular breathing such as shown in Additional file
1: Figure S4, deteriorates 4DMRI image quality, and potentially gives rise to unnecessary large treatment volumes when (unrepresentative) outlier tumor positions due to hiccups or a single deep inspiration are included. Therefore, not applying outlier rejection is not the optimal reconstruction strategy for both image quality and target definition. The dosimetric and clinical effect of excluding 5% of the data needs to be investigated and is a study on its own, as we focused on the quality of the reconstructed images. However, not covering the target for a certain part of the time during treatment in order to spare a large volume of healthy tissue is an accepted paradigm in radiotherapy as it is the base of treatment margin recipes [
31].
For radiotherapy treatment planning purposes, a 4DMRI should be precise, artifact-free and the reconstructed respiratory motion amplitude should be representative for the motion during the scan. The quality of 4DMRI datasets as measured by the set of quality parameters was significantly higher for the novel Min95 binning strategy compared to phase binning and the two other amplitude binning strategies. 4DMRI reconstruction was precise with a mean IBV of 1.6 mm, had smooth diaphragm profiles (S of 0.90), had a small inclusion range and did not suffer much from missing slices (RC of 95.5). It performed as good as applying strict outlier reduction (MeanIE) without the severe drawback of heavily underestimating the organ motion [
21]. Furthermore, this improvement was present for all subjects and the variation between subjects was reduced as well, indicating that the novel amplitude binning technique is robust against variation in patterns of irregular breathing. Each strategy has its strong points and weak points, e.g. a method that does not reject any outliers will result in a 100% DI although consequently it will suffer from artifacts resulting from low RC or high IBV, since in the case of irregular breathing, the chance that during sorting bin-slice combinations will not be filled increases (See Additional file
1: Figure S1). Choosing the optimal strategy will therefore be a well-argued balance between the performance of the various quality parameters.
4DMRI shows the tumor position in various respiratory phases. This can be integrated in a radiotherapy procedure similar to how 4DCT is used for lung and esophageal cancer treatment [
32,
33]. In case this 4DMRI technique is used for target volume delineation, the underestimation of the reconstructed motion should be taken into account. How to incorporate this into an ITV, or in a PTV margin in case a mid-position MRI is reconstructed from the 4DMRI, needs to be investigated in a future study. Furthermore, the described field of view is limited and needs to be expanded to encompass larger anatomical structures. This comes at a cost of scanning time. Reducing scanning time by acquiring fewer dynamics will potentially increase the incidence of missing bin-slice combinations and introduce a lower RC. A limitation of our reconstruction strategy is that we do not fill up missing bin-slice combinations when the RC is not 100%. A high RC still results in clinically usable 4DMRI since all anatomical structures and motion are still properly reconstructed. When the RC drops below 90% and the missing slices are in unfortunate positions (e.g., adjacent slices in the end-inhale and end-exhale bins), the quality of the 4DMRI deteriorates. A post-processing step such as interpolation of adjacent slices or adjacent respiratory bins might fill the missing bin-slice combinations, enabling for speeding up the image acquisition by lowering the number of dynamics [
30,
34]. However, a high RC will be preferable to such post-processing steps.
Acquiring a dataset pre-treatment for treatment planning purposes does not guarantee that the same (irregular) breathing motion will be present during the multiple radiotherapy treatment sessions. It is known that abdominal organ motion measured on pre-treatment imaging can be different from motion observed during treatment sessions [
4]. This is independent of the imaging technique, and as such present for 4DCT as well as in current clinical practice [
35]. However, compared to 4DCT the presented 4DMRI is acquired over an extended time of 6 min, increasing the possibility that the reconstructed motion of 4DMRI after outlier rejection is more representative than the motion observed with 4DCT which would take 2 min over such a field of view.