Supine MRI for regional breast radiotherapy: imaging axillary lymph nodes before and after sentinel-node biopsy

Regional radiotherapy (RT) is increasingly used in breast cancer treatment. Conventionally, computed tomography (CT) is performed for RT planning. Lymph node (LN) target levels are delineated according to anatomical boundaries. Magnetic resonance imaging (MRI) could enable individual LN delineation. The purpose was to evaluate the applicability of MRI for LN detection in supine treatment position, before and after sentinel-node biopsy (SNB). Twenty-three female breast cancer patients (cTis-3N0M0) underwent 1.5 T MRI, before and after SNB, in addition to CT. Endurance for MRI was monitored. Axillary levels were delineated. LNs were identified and delineated on MRI from before and after SNB, and on CT, and compared by Wilcoxon signed-rank tests. LN locations and LN-based volumes were related to axillary delineations and associated volumes. Although postoperative effects were visible, LN numbers on postoperative MRI (median 26 LNs) were highly reproducible compared to preoperative MRI when adding excised sentinel nodes, and higher than on CT (median 11, p  <  0.001). LN-based volumes were considerably smaller than respective axillary levels. Supine MRI of LNs is feasible and reproducible before and after SNB. This may lead to more accurate RT target definition compared to CT, with potentially lower toxicity. With the MRI techniques described here, initiation of novel MRI-guided RT strategies aiming at individual LNs could be possible.


Introduction
In recent years, regional treatment for breast cancer patients with tumour-positive sentinel nodes (SNs) has changed. Regional radiotherapy (RT), performed after breast-conserving surgery (BCS) or mastectomy, is increasingly used as an alternative to axillary lymph node (LN) dissection (Giuliano et al 2010, Galimberti et al 2013, Donker et al 2014. In conventional regional breast RT, target volumes are contoured on computed tomography (CT) scans, according to contouring guidelines, including those recently published by the European Society for Radiotherapy and Oncology (ESTRO) (Offersen et al 2015). These guidelines focus on the delineation of axillary levels based on vessels and muscles, since LN visibility on CT is limited. Furthermore, the location of regional LNs varies greatly between patients (Dijkema et al 2004). Magnetic resonance imaging (MRI) can image individual axillary LNs better (Lernevall 2000, Harnan et al 2011, Rahbar et al 2012, Ahmed et al 2014, Li et al 2014, Menezes et al 2014, Kuijs et al 2015 in direct relation to surrounding anatomical structures, e.g. the brachial plexus, heart, and chest wall. Anatomical information on exact LN locations obtained from dedicated MRI techniques may lead to adapted regional breast RT target volumes, such that more healthy tissue can be spared. This could result in decreased RT-related toxicity, consisting of arm morbidity, e.g. pain, edema, and shoulder stiffness. To our knowledge, this has not been investigated previously.
Conventional diagnostic breast MRI is not optimized for RT planning, and is acquired in prone position. We have previously developed dedicated MRI, in healthy volunteers, to image axillary LNs for RT planning in supine treatment position (van Heijst et al 2016). Both anatomy and posture endurance in patients undergoing MRI are likely to differ from that of healthy volunteers (Luscieti et al 1980, Hadamitzky et al 2010. The next step is, therefore, to validate MRI for regional RT in breast cancer patients who are scanned in a treatment position. In standard elective RT, entire axillary LN levels are irradiated. Using MRI, LNs could potentially be delineated and targeted individually. This makes the reproducibility of detecting LNs of the utmost importance. Furthermore, LN identification on MRI may be affected by axillary surgery. This is usually limited to sentinel-node biopsy (SNB), which is performed for staging purposes, in combination with breast surgery. The surgical excision of one or more SNs in SNB leads to anatomical alterations, potentially also due to inflammation and seroma formation (Srivastava et al 2012). Moreover, patient posture endurance for undergoing MRI may be affected by surgery.
The purpose of this work is to test the reproducibility of LN detection using MRI, and posture endurance, in breast cancer patients who undergo SNB, pre-and postoperatively, in supine treatment position. Furthermore, anatomical locations and volumes of all LNs, including SNs excised during SNB, are related to the standard delineation of axillary levels. LN detection rates of MRI are compared to standard postoperative CT.

Patients
The Internal Review Board of University Medical Center (UMC) Utrecht approved the study, which is registered under Netherlands trial number NL50046.041.14. Written informed consent was obtained from 23 female patients (table 1), who were diagnosed with either ductal carcinoma in situ (DCIS), or early-stage breast cancer without pathologically involved LNs or distant metastasis (cTis-T3N0M0). They were scheduled for SNB, performed during BCS or mastectomy, at UMC Utrecht or at Diakonessenhuis Hospital, in Utrecht, The Netherlands. No surgery or RT had been performed previously in the ipsilateral axilla. One patient underwent mastectomy, while 22 participants underwent BCS, 21 of whom received subsequent standard RT treatment at UMC Utrecht.

MRI set-up
Scanning was performed on a 1.5 T wide-bore MRI scanner (Philips Ingenia, The Netherlands) at the RT Department of UMC Utrecht, from February 2015 until February 2016. Participants were scanned in supine RT position (van Heijst et al 2016), i.e. with arms in abduction, supported by a wedge-board (Thorawedge, CIVCO Medical Solutions, The Netherlands). An anterior receive coil was connected and placed on an adjustable polymethyl-metacrylate support, such that the outer anatomy was not deformed. A posterior receive coil was located in the scanner table.

MRI and CT scanning
Participants underwent two MRI scan sessions. The first was performed a median 7 d (range: 1-23 d) before surgery, whereas the second took place a median 20 d (range: 8-45 d) after surgery. Five MRI scans were acquired on both occasions, which were optimized previously in healthy volunteers (van Heijst et al 2016). Relevant imaging parameters are listed in table 2. No contrast agent was administered.
The first method was a T1-weighted (T1w) spoiled gradient echo (SPGR) technique, acquired transversally. The in-plane field of view (FOV) encompassed the outer body contour, ranging in the cranial-caudal direction from the mandible to the caudal border of the breast. Here, multi-point Dixon (mDixon) water-fat separation was applied, where two echoes are used to separate signal from fat and water (Eggers et al 2011). The other four MRI scans were acquired in the coronal plane, with FOVs covering the axillary areas. One MRI technique was a T1w SPGR sequence using mDixon, with a smaller FOV than the first T1w SPGR, but higher in resolution. Furthermore, two-dimensional (2D) and 3D T2w fast spin echo techniques were both applied. The final scan was a diffusion-weighted MRI (DWI), in which restriction of water mobility is reflected by hyper-intense signal.
Twenty-one patients underwent postoperative CT imaging for standard RT planning in supine RT position, on a support (CIVCO, The Netherlands). This was usually performed on the same day as postoperative MRI. Acquisition duration was 20 s, slice thickness 3 mm, and minimal in-plane resolution was 1 × 1 mm 2 (Brilliance, Philips). The FOV of the RT-planning CT included the breast and axillary regions.

Posture endurance for MRI
Acquisition time of MRI was limited to 20 min per session. With the additional time needed for positioning and scanner preparation, the supine RT position had to be maintained for up to 30 min in each session. Since this is potentially difficult to endure, we monitored whether the patient was able to finish the procedure. If applicable, the reason for the insufficient posture endurance was recorded, as was the need for any adjustments in positioning.

Contouring of axillary levels on MRI
The first T1w scan, with the largest FOV, was considered as the reference scan, to which other scans of the same MRI session were co-registered. This was performed by a rigid registration, based on anatomical reference structures (e.g. large blood vessels or muscles) which were not close to the SNB location. Since automatic cross-contrast registration algorithms are complex and prone to error, this was performed manually. For MRI methods using mDixon, water-only (Dixon-water) and fat-only (Dixon-fat) images were both examined. A breast radiation oncologist contoured standard regional target volumes on preoperative scans (figure 1), according to ESTRO delineation guidelines (Offersen et al 2015). These were developed for CT, but are   also applicable for use on MRI. For MRI, minor adaptations were implemented due to softtissue differences. We will refer to axillary target contours, defined as levels I-IV, as axillary levels. The interpectoral level was incorporated in level II.

LN identification and delineation
Using information from all scans in one MRI session, individual LNs were identified, and delineated on the reference scan (figure 1). LN identification was first performed by two observers, who were trained for this by a breast radiologist, and who established an intermediate consensus on structure identifications by using pre-and post-SNB scans. Identification on MRI and CT was validated by two other observers-a breast radiation oncologist and a breast radiologist. After multidisciplinary meetings with all observers, final consensus was reached on pre-and postoperative sets of LNs, on both MRI and CT.
The number of LNs was determined for each patient in scans from both MRI sessions. Moreover, it was determined in which axillary level a LN was located on the preoperative MRI. If an LN was partly located in an axillary level, it was considered to be included in that contour. If an LN was located in more than one level, the level was chosen in which it occupied the greatest volume. Finally, LNs outside the axillary level contours, but within 5 mm, were counted separately. LNs at more than 5 mm distance were excluded. Pre-and postoperative sets of delineated LNs in each patient were compared on a node-to-node basis. If possible, spatial correlation was established between individual LNs on both MRI sessions.
Furthermore, if possible, the excised SNs were retrospectively identified on the preoperative MRI. This was performed visually, by comparing surrounding LNs and studying anatomical changes. Moreover, the number of excised SNs was taken from the pathology report and compared to differences in the number of LNs before and after SNB. During SNB, SNs excised can also be internal mammary LNs. However, for our analysis we consider only axillary SNs.
The LN detection rate is defined by the number of identified LNs. These were compared per patient, on MRI before and after SNB, while also accounting for the excised SN(s). The postoperative LN detection rate on CT was compared to the postoperative MRI detection rate.

Statistical analysis
Statistical analysis was performed using R software (R Core Team 2014). Normality was checked by visual inspection of calculated Q-Q plots. To assess reproducibility, the LN detection rate on pre-SNB MRI was compared with the LN detection rate on post-SNB MRI using paired Wilcoxon signed-rank tests. Similarly, the LN detection rate of post-SNB MRI was compared to post-SNB CT. Furthermore, Bland-Altman plots were calculated, which represent the difference between the LN detection rates as a function of the average of the two measurements.

Volumes of LNs and axillary levels
On MRI, the summed volumes of all delineated LNs were determined, in order to indicate potential reduction in target volumes for elective RT, if based on individual LNs instead of entire standard axillary levels. LN-based volumes were compared to volumes of the respective axillary levels. The effect of expanding the LN delineations was also explored by making an isotropic expansion of 5 mm for each individual LN delineation, and performing a similar comparison of the resulting encompassing volume, compared to the axillary levels.

Posture endurance for MRI
Twenty out of 23 patients completed both scan sessions without posture adjustments. One patient experienced pain in her shoulders during both pre-and postoperative MRI. As she was unable to finish either session, not all scans were acquired. Another patient experienced pain in the ipsilateral shoulder, preoperatively, due to a pre-existing cervical hernia. A third patient had a painful contralateral shoulder during postoperative MRI. Patient positioning was adjusted for both of these patients after acquisition of three scans, by lowering the ipsilateral or contralateral arm. The remaining scans were acquired without problems.

Use of MRI in practice
LNs were visualized in all patients (figures 1 and 2). Fat separation by mDixon was reliable. For LN identification, all acquired MRI scans could be used. In practice not all scans were always contributive. Most LNs were first identified on the T1w SPGR reference scan. The T2w TSE scans-both showing similar contrast-were consulted to confirm this, and to identify additional LNs, usually close to blood vessels and arteries, due to lack of contrast on T1w images in those locations. A drawback was that image quality of the T2w scans was lower in patients with higher BMI. The other T1w method showed similar contrast as the reference scan and therefore served as confirmation for LN identification. The added value of DWI, while showing some LNs in high contrast, was limited due to low resolution.

Postoperative artifacts
In all patients, postoperative effects were recognizable on MRI. Enlargement of LNs was observed after SNB in some patients. Repositioning of arteries, muscles and LNs was observed in all patients. Spatial differences related to this were usually small. Formation of postoperative seroma was also visible in all patients, which appeared diluted in all but two patients. The presence of post-SNB effects served as landmarks for retrospectively localizing SNs on preoperative MRI that were excised during SNB (figure 3).

MRI: detection rates, spatial correlation, and SNs
The median number of LNs was 26 (range: 14-42) on preoperative MRI, and was 25 (range: 13-40) on postoperative MRI (table 3). In total, 617 and 587 LNs were identified, on pre-and post-SNB MRI, respectively. The median number of SNs excised during SNB was 1 (range: 1-3), with a total of 33. For 22 out of 23 patients, the preoperative LN detection rate was equal to the postoperative detection rate on MRI, when the SNs were added (p = 1.00) (table 3, figure 4). In one patient, 20 LNs were identified before SNB and 17 after, while 2 SNs had been removed.
In 20 out of the remaining 22 patients, all 550 individual LNs could be spatially correlated between MRI sessions. Any anatomical variation due to a different scanning position did not affect LN correlation in general. In the two other patients, a large postoperative seroma, as well as shifted muscles and arteries, disrupted the anatomy, thereby hampering an exact correlation. However, pre-and postoperative MRI detection rates were equal.
All SNs were identified retrospectively on preoperative MRI. For all patients, numbers of identified SNs matched numbers of excised SNs described in pathology reports (table 3).

Detection rate on postoperative CT
In total, 255 LNs were identified on all 21 postoperative CT scans. The median number of LNs was 11 (range: 7-22) per patient. LN detection rate on CT was significantly lower than on postoperative MRI (p < 0.0001) (table 3, figure 4).

Relative location of LNs to axillary levels
The majority (353, 57%) of all 617 identified LNs were located in axillary level I (median per patient: 15, range: 13-28) (figure 5). In particular, all 33 identified SNs were identified in level I. The more cranially located an axillary level was, the fewer LNs it tended to contain. A total of 138 (median per patient 6, range: 4-12) LNs were in level II. Only one patient showed more LNs in level II than in level I. Summed over all patients, levels III and IV contained only 42 and 28 LNs respectively (7% and 5%), at a median of 1 LN per patient (ranges: 0-4 and 0-3, respectively). Moreover, 69 LNs (11%) were counted outside the axillary levels, at a median of 2 per patient (range: 0-8). Most of these LNs (49 of 69, 71%) were located nearest to level I.

LN-based volumes
The summed volume of LNs in level I was the largest, in all patients (table 4). Median LN-based volumes were at most 3% of the volumes of the respective axillary levels. After the 5 mm expansion of each individual LN delineation, the encompassing volume was limited to at most 46% of the corresponding axillary level. The median encompassing volume of LNs counted outside axillary levels on MRI was 5.0 cm 3 after the 5 mm expansion.

Discussion
MRI of axillary LNs, in supine RT position, is feasible, before and after SNB. Although diagnostic MRI for axillary imaging is thoroughly studied (Lernevall 2000, Harnan et al 2011, Rahbar et al 2012, Ahmed et al 2014, Li et al 2014, Menezes et al 2014, Kuijs et al 2015, MRI studies for target delineation in regional breast RT are scarce. SNB did not affect patient posture endurance for MRI in supine position. LN detection on MRI scans before and after SNB was highly reproducible, despite postoperative artifacts observed in all patients. All SNs that were excised during SNB were retrospectively identified. Moreover, in 20 patients, all other LNs could be individually correlated between MRI sessions. Compared to MRI, the LN detection rate on CT was significantly lower. Note: LN = lymph node, preop = preoperative, MRI = magnetic resonance imaging, postop = postoperative, CT = computed tomography, ΔLNs = difference in numbers of lymph nodes, SN = sentinel node, PA = pathology, NA = not applicable (since no CT was performed for those patients). a In patients 5 and 14, one SN was an internal mammary LN; it was therefore not counted on preoperative MRI in our analysis.
There is no gold standard available for the number of LNs. The numbers in our study, ranging from 14 to 42 on MRI, fall within the expected range of 6-45 LNs, based on previous imaging (MRI, CT, and ultrasonography) and histological studies on LNs (Jiang et al 2007, Luciani et al 2009, MacDonald et al 2010, Korteweg et al 2011, Nakai et al 2011, Hong et al 2012. Histological information, after axillary dissection, is limited to axillary levels I and II, while our work aimed to image LNs in axillary levels I-IV. The range is also similar to that in healthy volunteers (van Heijst et al 2016). Most LNs (57%) were located in level I, being the largest axillary volume, including all identified SNs. Two SNs were tumour-positive, but these  were not obviously enlarged on MRI. We did not perform a diagnostic analysis on MRI for the SNs. The more cranially located the level was, the fewer LNs it tended to contain. Numbers of LNs detected in levels III and IV were within the expected range of 1-10 LNs (Ying and Ahuja 2003). However, the supraclavicular region was difficult to image, due to signal from large blood vessels that was not always adequately suppressed in T2w scans, and which could obscure other structures in T1w images. This could have impaired localization of small LNs, possibly hampering the consistency of LN detection on MRI.
The LN detection rate on MRI was significantly higher than on CT, mainly due to its superior soft-tissue contrast. It should be noted that CT was acquired for standard RT target delineation and was not optimized for LN identification.
There are limitations in the set-up of our study. Consensus was reached on LN identification with the expertise of a multidisciplinary team of observers. This did not allow determination of inter-observer variation, as LN identification was performed in consensus, by comparing pre-and postoperative scans together, and in a relatively limited number of patients. Most discussion was on LNs that were located close to blood vessels, the signal of which was not suppressed adequately. This might have led to false positives and false negatives. The current work focused on establishing whether it is feasible to reproduce LN detection before and after SNB, where information from preoperative scans was used in postoperative scans, and vice versa. An inter-observer delineation study will be conducted in the future to assess the potential bias introduced this way. Confirmation of the results in our current study will be investigated in a larger group of patients.
Using ESTRO guidelines (Offersen et al 2015) on MRI, up to 8 LNs per patient were located outside the axillary level delineations (but within 5 mm distance), commonly cranial to levels III/IV or dorsolateral to level I. MacDonald et al (2010) also found LNs outside axillary levels. The question is whether these LNs should be irradiated. One approach to cover these in the target is to extend the axillary contours (MacDonald et al 2010), leading to more treated normal tissue. Indication for current elective RT treatments is determined by results from clinical studies that have investigated anatomical locations of regional recurrences (Offersen et al 2013). Delineation guidelines used for regional RT are based on these findings (Nielsen Table 4. LN-based target volumes, without and with expansion, per respective axillary level. Volumes based on individual LNs as delineated on MRI are calculated both without expansion (summed volume), and with a 5 mm expansion of each individual LN delineation (encompassing volume). Volumes of respective axillary levels (I-IV) are given, delineated on MRI according to modified ESTRO delineation guidelines (Offersen et al 2015). LNs counted outside of the levels are indicated as 'other'. Values are denoted 'median (range)'. Values < 0.05 cm 3 were rounded up to 0.1 cm 3 . LN = lymph node.  Offersen et al 2013, Gentile et al 2015, Verhoeven et al 2015. One could therefore argue that LNs outside the elective axillary levels are not clinically relevant. The addition of MRI information to standard CT-based RT planning could lead to a more accurate definition of target volumes, by localizing individual LNs. Therefore, we explored individual LN target delineations and calculated LN-based volumes. Even with a 5 mm expansion for each LN delineation, the encompassing volumes are considerably smaller than the respective axillary levels, which are delineated based on vessels and musculature. This signifies that, for elective CT-based RT strategies, adding information from dedicated MRI scans could lead to smaller treated volumes, while still effectively covering the individual LNs.
To implement this into the current clinical CT-based workflow, MRI could be co-registered with the RT-planning CT scan, which could induce set-up variations and registration errors. It is important to perform both scans in the same patient position as much as possible, in order to keep anatomical differences small. Furthermore, the interval between acquisition of the pre-RT MRI and RT-planning CT scans should be as short as possible. Uncertainties due to motion between RT fractions should be minimized. A suitable method could be to use the same dedicated (MRI compatible) positioning device on both CT and MRI modalities, and on the linear accelerator. The impact of any remaining errors due to inter-fraction motion should be quantified. However, the potential clinical benefit of adding MRI in the current clinical process is high.
As shown in our study, significantly more LNs are found on MRI than on CT. Moreover, they are imaged in a direct relationship to surrounding radiosensitive structures, e.g. the heart, brachial plexus, and chest wall. The RT planning volume could be adapted accordingly by restricting the RT dose more conformally to the LNs, thereby sparing healthy tissue, decreasing the overall treated volume, and potentially reducing toxicity. Furthermore, nowadays, elective RT of LNs is performed after axillary surgery. Since the presence of postoperative artifacts could obscure LNs, new approaches may focus on preoperative RT instead.
On-line MRI guidance, as achieved by the UMC Utrecht MRI/linear accelerator and other MRI/RT modalities (Dempsey et al 2006, Lagendijk et al 2008, Fallone et al 2009, Raaymakers et al 2009 could be beneficial for LN-based RT treatments by reducing set-up uncertainties, as ideally this is performed in an MRI-only workflow. Moreover, the clinical availability of these hybrid MRI/RT modalities could open further possibilities for treatment approaches of individual LNs, such as the delivery of a stereotactic high dose to tumourpositive LNs. For such advanced stereotactic RT techniques, the use of additional methods for identifying metastatic LNs on MRI is necessary (Harnan et al 2011). Such novel approaches may lead to lower RT-related toxicity, shorter RT treatment schedules, or omission of surgery for selected breast cancer patients. The feasibility of MRI-guided RT, specifically targeting individual LNs, is currently under investigation at UMC Utrecht.

Conclusion
MRI of LNs in supine RT position in breast cancer patients is feasible. The LN detection rate on MRI is reproducible before and after SNB. Implementation of MRI in RT planning may lead to more accurate target definition, compared to CT-based regional RT planning. LN-based delineation using MRI may result in a reduction in standard regional RT target volumes. This may facilitate new regional RT approaches in breast cancer patients targeting individual LNs.