Network meta‐analysis of novel and conventional sentinel lymph node biopsy techniques in breast cancer

Background The aim of this network meta‐analysis was to compare the performance of blue dye alone or in combination with radioisotope (technetium‐99m, Tc) with three novel techniques for sentinel lymph node detection in breast cancer: indocyanine green fluorescence (ICG), superparamagnetic iron oxide (SPIO) nanoparticles and contrast‐enhanced ultrasound imaging (CEUS). Methods PubMed, Embase, the Cochrane Library, China Knowledge Research Integrated Database, ClinicalTrials.gov and OpenGrey databases were searched up to 31 November 2017, without language restriction. Studies that compared the detection performance of at least one of the novel methods (ICG, SPIO and CEUS) with that of traditional methods (blue dye and/or radioisotope) were included in network meta‐analysis. Results Thirty‐five studies were included. Pooled risk ratios (RRs) for Tc (1·09, 95 per cent c.i. 1·04 to 1·15), ICG (1·12, 1·07 to 1·16) and SPIO (1·09, 1·01 to 1·18) showed statistically better performance in detecting sentinel lymph nodes than blue dye alone. ICG had the lowest false‐negative rate, with a RR of 0·29 (0·16 to 0·54), followed by Tc (RR 0·44, 0·20 to 0·96) and SPIO (RR 0·45, 0·14 to 1·45), with blue dye alone as the reference group. Conclusion SPIO or ICG alone are superior to blue dye alone and comparable to the standard dual‐modality technique of blue dye with Tc.


Introduction
Lymphatic staging of breast cancer was initially established by axillary lymph node dissection followed by sentinel lymph node biopsy 1 . There may be more than one sentinel node, depending on the location of the tumour and the network of surrounding lymph vessels 2 . For patients with early breast cancer and clinical and radiological absence of axillary lymph node metastases, sentinel lymph node biopsy is currently considered the standard of care 3 .
The most widely used technique for sentinel lymph node identification is the dual-modality method involving the injection of technetium-99m (Tc)-labelled nanocolloid and blue dye (BD) into the peritumoral or periareolar region 4 . 'Hot' nodes detected using a handheld scintillation counter (γ-probe) and 'blue' nodes identified during surgery are both recognized as sentinel nodes and removed, followed by histopathological confirmation of metastases. An identification rate of 96 per cent was reported in a large meta-analysis 5 of 8000 patients. In another meta-analysis 6 of more than 9000 patients, the false-negative rate was 5⋅9 per cent with the dual-modality technique, in comparison with 8⋅6 and 7⋅4 per cent for dye-only and tracer-only techniques respectively 6 . The dual technique, nevertheless, has shortcomings, including radiation exposure to healthcare professionals and patients, issues with radiotracer availability, dependency on availability of nuclear medicine units and allergic reactions to BD.
New techniques have been developed to improve the clinical value of sentinel lymph node biopsy with similar accuracy, but avoiding irradiation and risks of allergy. Novel techniques studied in recent years include those using indocyanine green (ICG) fluorescence, superparamagnetic iron oxide (SPIO) nanoparticles, and contrast-enhanced ultrasound imaging using microbubbles (CEUS). A systematic review by Ahmed and colleagues 7 in 2014 reported that these techniques showed clinical potential but had high false-negative rates. The review was based on pairwise meta-analysis and was unable to compare relative performance between techniques not previously studied in a direct comparison. The aim of the present network meta-analysis was to compare these techniques directly and indirectly in terms of detection and false-negative rates.

Literature search and eligibility criteria
A literature search was performed in electronic databases, including PubMed, Embase, the Cochrane Library, China Knowledge Research Integrated Database, ClinicalTrials .gov and OpenGrey, up to 31 November 2017. The specific concepts used in the search strategy were 'Breast Neoplasms', 'Sentinel Lymph Node Biopsy', 'Indocyanine Green', 'Superparamagnetic Iron Oxide' and 'Contrast-enhanced Ultrasound'. Medical Subject Headings (MeSH) or Emtree, and free-text terms were both searched (Table S1, supporting information). There were no restrictions on language. All bibliographies listed in relevant review papers and included publications were also checked.
Two reviewers independently searched for eligible studies based on predefined eligibility criteria. Studies that compared the detection performance of at least one of the novel methods (ICG, SPIO and CEUS) with that of traditional methods (BD and/or Tc) were included. Single-arm studies that assessed only one of the methods, reviews, case reports, animal studies, commentaries and letters to editors were excluded. Discrepancies were discussed between reviewers until consensus was reached.

Data extraction and quality assessment
The following data were extracted from the included studies: study characteristics (publication year, number of subjects); baseline characteristics (mean age and recruitment method); and outcome events (number of detections and number of false-negative reports). Relevant data were also extracted for quality assessment.
A network geometry was constructed based on the included studies 8 . Each node represented various techniques, and the size was weighted by the number of subjects from whom data on each technique were derived. The connecting line between two nodes represented the presence of a direct comparison, and the number of studies included determined the thickness of each line. Hence, the thicker the line, the more studies comparing the two methods were included.
The quality of each study was evaluated by two independent investigators, using the Risk Of Bias in Non-randomized Studies of Interventions (ROBINS-I) tool 9 . Seven domains (confounding, selection of participants into the study, classification of interventions, deviations from intended interventions, missing data, measurement of outcomes and selection of the reported result) were evaluated to assess the risk of bias. For each of the seven domains in the ROBINS-I tool, responses to the signalling questions were taken together to arrive at the judgement of low, moderate, serious or critical risk of bias, and judgements within each domain were summarized to provide an overall risk-of-bias judgement for the outcome being assessed. Any disagreement in quality assessment was resolved by discussion and consensus.

Statistical analysis
A network meta-analysis, comparing the detection methods for the proposed outcomes, was performed using a multivariable meta-regression model with random effects, adopting a frequentist approach. The model allowed the inclusion of potential co-variables and accounted for correlations from multiarm trials. Risk ratios (RRs) with 95 per cent confidence intervals were estimated for detection rate and false-negative rate.
To rank the effect size of all the methods, surface under the cumulative ranking (SUCRA) value, indicating the probability of each method being the best, was used 10 . A higher SUCRA value indicated a higher rank of performance for the method. To check for the assumption of consistency, the design-by-treatment approach was used. This approach tested whether the difference between RRs from direct comparison (either by 2-arm or multiarm design) and those from network meta-analysis was statistically significant. The χ 2 test was used to test the difference throughout the entire network, and P < 0⋅050 indicated the need for further investigation to identify the sources of inconsistency 11 .
The network meta-analyses were generated using Stata/MP ® version 13 with network and network graphs packages 12 -14 .
Quality evaluation of all the studies demonstrated a low to moderate risk of bias in the seven domains assessed, and 20 of 35 had a moderate risk of bias on overall assessment ( Table S2, supporting information). Moderate risk of bias was assigned in terms of patient selection to studies where the details of recruitment were not clearly reported.

False-negative rate
The SUCRA results for false-negative rate showed a similar pattern to those for detection rate; CEUS and BD had a low SUCRA rank of 6 and 5 respectively, which translates to higher false-negative rate compared with other modalities ( Table 4).

Network consistency
Forest plots comparing pooled estimates from both direct comparison and network meta-analysis were reported for detection rate (Fig. S1, supporting information) and false-negative rate (Fig. S2, supporting information). On visual check, greater consistency existed among pooled estimates from all two-arm studies, all multiarm studies and network meta-analysis. The consistency assumption was examined statistically using the χ 2 test,  SUCRA, surface under the cumulative ranking; ICG, indocyanine green; SPIO, superparamagnetic iron oxide; Tc, technetium-99m; Tc/BD, combined use of technetium-99m and blue dye; CEUS, contrast-enhanced ultrasound imaging; BD, blue dye alone.
which yielded P = 0⋅316 for detection rate and P = 0⋅948 for false-negative rate, indicating no evidence for the potential violation of assumption. There was, therefore, no statistically significant evidence of inconsistency.

Discussion
In this review, the ICG fluorescence technique ranked first in terms of detection, as well as having the lowest false-negative rate of the techniques compared. The ranking of other techniques in descending order was SPIO, Tc, dual modality (Tc/BD), CEUS and BD in terms of detection rate. The use of ICG fluorescence for sentinel lymph node biopsy in patients with breast cancer was first described in 1999 by Motomura and colleagues 50 , who reported an identification rate of 73⋅8 per cent. The main advantages of this method are accurate placement of the skin incision as subcutaneous lymphatic channels can be identified on the skin before operation, and accuracy is improved with the use of fluorescence as there is high sensitivity with illumination even if the nodes do not stain green. On the contrary, problems and technical difficulties associated with this technique include difficulty with ICG fluorescence detection for lymph nodes more than 1 cm below the skin, need for operating theatre lights to be turned off during fluorescence navigation, and diffusion of ICG into subcutaneous tissue if the procedure takes more than 30 min, reducing the sensitivity of fluorescence detection. In a 2016 meta-analysis 51 , the pooled detection rate for this technique was 0⋅98 (95 per cent c.i. 0⋅96 to 0⋅99) with a false-negative rate of 8 per cent. The main limitation of the meta-analysis was that only six studies were included in the final analysis. The present meta-analysis included 22 studies that used ICG. This technique achieved a significant 12 per cent improvement in terms of detection rate and a 71 per cent lower false-negative rate compared with BD.
SPIO nanoparticles (Sienna+ ® ; Sysmex Europe, Hamburg, Germany) have gained interest since the first description of their use in 2013 52 . SPIO nanoparticles stay in the tissue for a prolonged period and can be detected with a magnetometer 25 . This characteristic has made it feasible for SPIO to be injected in the outpatient clinic 3-15 days before surgery, with a success rate comparable to that of Tc/BD 53 . A meta-analysis 54 of seven randomized trials in 2016 showed that the SPIO method was not inferior to the standard technique in identification rate (97⋅1 versus 96⋅8 per cent), total lymph nodes retrieved (1⋅9 versus 1⋅8 nodes per patient) or false-negative rate (8⋅4 versus 10⋅9 per cent). Using both direct and indirect comparisons, the present analysis found that SPIO performed significantly better than BD in terms of detection rate. Although the RR compared with BD for false-negative rate was not statistically significant, a 55 per cent decrease on the point estimate suggested a promising direction for the use of SPIO in sentinel lymph node biopsy.
CEUS with microbubbles, based on the use of dispersion with sulphur hexafluoride gas injected intradermally around the areola, for sentinel node biopsy was first reported in 2013 55 . In a subsequent meta-analysis of five studies 56 , this technique identified sentinel lymph nodes in 9⋅3-55⋅2 per cent of patients, with a sensitivity of 61-89 per cent and a false-negative rate of 6⋅6-39 per cent. No comparative study of CEUS against the dual-modality technique exists to validate this method. In the present analysis, there was no statistical difference between CEUS and BD in terms of detection rate, but the risk of a false-negative result was four times that for the Tc method (RR 4⋅00, 95 per cent c.i. 1⋅17 to 13⋅70). The high false-negative rate might be attributable to inconsistent identification of sentinel lymph nodes owing to interobserver variability, even when the same ultrasound equipment is used. The introduction of super-resolution imaging, ultrafast ultrasound imaging and improved microbubble transit may improve the visualization of lymphatics and sentinel lymph nodes with this technique 55 .
It is acknowledged that this review has inherent limitations as evaluation of the performance of each technique was limited to detection and false-negative rates. From a clinical standpoint, the strength of a particular modality is also dependent on other factors such as ease of use and side-effects. It would be of interest to investigate other details pertaining to the use of these techniques, such as dose and site of injection, to see how these affect performance. Limitations in terms of statistical analysis related to the lack of RCTs, although the risk of bias was generally low to moderate among the included studies.
The strength of the present review is the use of network meta-analysis to allow direct and indirect comparisons between techniques for sentinel lymph node identification and false-negative rates. Previous systematic reviews evaluated various techniques qualitatively, rather than quantitatively, making conclusions less convincing. By synthesizing direct and indirect evidence using network meta-analysis, this study was able to rank the detection and false-negative rates across all techniques quantitatively.
Both ICG and SPIO consistently performed better than BD alone in this review. Based on this finding, these methods could provide a reliable alternative in centres where BD is used as a single modality, as the use of either technique would result in better detection and false-negative rates. Although it is interesting to note that these two novel methods had similar performance to radioactive colloid in terms of detection as well as false-negative rates, technical, logistical and economic factors should be taken into consideration. Studies comparing novel methods with standard dual-modality techniques are still needed, in the hope of demonstrating superior performance.
In this network meta-analysis, Tc, ICG and SPIO demonstrated superior performance over BD in terms of detection and false-negative rate. Future cost-effectiveness analyses will be crucial to determine whether the novel techniques are sustainable in the long run.

Disclosure
The authors declare no conflict of interest.