The predictive value of lymph node status pre-operation by ultrasound and MRI in early and newly diagnosed breast cancer

Abstract

Background

Accurate pre-operative assessment of axillary lymph node status is key for planning surgeries in early and newly diagnosed breast cancer. Ultrasound (US) and magnetic resonance imaging (MRI) are widely used, but their performance is inconsistent across studies. This meta-analysis evaluates and compares the predictive performance of MRI and US for pre-operative lymph node metastasis in early and newly diagnosed breast cancer.

Methods

We searched PubMed/MEDLINE, Embase, Scopus, and Web of Science (inception-31 December 2025) for studies on diagnostic performance in preoperative axillary staging using ultrasound and/or MRI. Two independent reviewers extracted or reconstructed true positive, true negative, false positive, false negative, and applied QUADAS-2. Sensitivity and specificity were pooled with random-effects GLMM, with direct comparisons and prevalence-based predictive values. 

Results

Twenty-nine studies (n = 8,544) were included. Pooled sensitivity/specificity were 61.1%/92.0% for MRI and 59.8%/90.3% for ultrasound, yielding LR + 7.64 vs. 6.19 and DOR 18.1 vs. 13.9. HSROC AUCs were 0.863 (MRI) and 0.843 (ultrasound). In paired cohorts (k = 12), sensitivity did not differ, and MRI showed a non-significant specificity advantage.

Conclusion

 MRI and ultrasound showed comparable, moderate sensitivity and high specificity for preoperative axillary metastasis. MRI tended toward higher specificity and positive predictive value, whereas sensitivity did not differ in paired studies. Imaging can guide triage, but histopathology remains necessary for definitive staging.

Supplementary Information

The online version contains supplementary material available at 10.1007/s10238-026-02087-6.

Introduction

Breast cancer is the most diagnosed malignancy among women worldwide and is a leading cause of cancer-related mortality even with continued advances in screening, molecular classification, and treatment strategies [1]. Recent global statistics show that over 2.3 million women are diagnosed with breast cancer each year, with more than 600,000 related mortalities [2]. Earlier detection and having a multidisciplinary approach have improved survival, but disease prognosis still varies widely because of different tumor biologies and disease progression at the presentation [3]. Axillary lymph node (ALN) status is a strong predictor of recurrence and overall survival in the breast cancer, therefore accurate assessment of ALN is important for determining stage, choosing best treatment strategy regarding surgical interventions, systemic treatment, and radiotherapy [4, 5]. Axillary lymph node dissection (ALND) has historically been the gold standard for evaluating nodal metastasis. Also, it is clearly associated with higher morbidity such as nerve injuries, restricted mobility of shoulder, more pain and infection, and lymphedema [6]. To minimize these complications, sentinel lymph node biopsy (SLNB) has become the current standard of care for clinically node-negative early breast cancer and povides accurate staging while significantly reducing surgical morbidity [7]. On the other hand, SLNB is still an invasive procedure and may not be necessary for all patients, especially for patients whose nodal involvement probability is low. Furthermore, a proportion of patients still undergo unnecessary nodal surgery due to limitations in preoperative clinical evaluation [6, 8].

These issues have led to the development of non-invasive imaging tools capable of effectively predicting lymph node status before surgery; these tools present the potential to stratify patients and guide surgical decisions before an incision is made. Among various modalities, ultrasonography (US) and magnetic resonance imaging (MRI) are the most widely employed tools to evaluate nodal involvement in a preoperative setting [9]. Ultrasound is frequently used owing to its ready availability, low cost, absence of radiation, and ability to assess morphological features of axillary lymph nodes such as shape, cortical thickness, and hilum appearance [10]. MRI, on the other hand, provides superior soft-tissue contrast and offers thorough visualization of axillary anatomy. Functional sequences, such as diffusion-weighted imaging, may further improve the detection of metastatic nodes [11].

Despite the documented utility of US and MRI, their diagnostic performance varies substantially across studies. Reported sensitivities of ultrasound range from 26% to more than 80%, reflecting differences in operator skill, criteria for positive nodes, and patient populations [12]. While MRI generally demonstrates higher sensitivity than US, its specificity is variable, and access, cost, and standardization challenges remain significant. A better understanding of their predictive capabilities may inform clinical pathways that minimize invasive procedures and improve patient outcomes [13].

The application of preoperative ALN status assessment has increased to lessen the axillary surgery extension in early breast cancer. Still, the performance of US and MRI widely varies in the literature. Prior meta-analyses have generally evaluated each modality in isolation, applied inconsistent positivity thresholds, or included few paired cohorts, making it difficult to quantify the clinical trade-off between false positives and false negatives. Moreover, most syntheses stop at sensitivity and specificity and do not translate accuracy into predictive values across plausible pre-test risks, which limits bedside interpretability. To address these gaps, we performed an updated diagnostic performance meta-analysis to compare ultrasound and MRI using a consistent modeling framework, incorporating head-to-head within-study contrasts when available, and estimating prevalence-dependent predictive values to support patient-specific decision-making. We further provide head-to-head comparisons, scenario-based PPV/NPV estimates, and an assessment of small-study effects to clarify the clinical implications and robustness of the pooled diagnostic accuracy estimates.

Methods

We undertook this systematic evidence review based on best practices and reported the findings as per a checklist provided by PRISMA (the Preferred Reporting Items for Systematic Reviews and Meta-Analyses) [14].

Eligibility criteria

We included original, human diagnostic accuracy studies that evaluated preoperative axillary nodal status using ultrasound and/or MRI in patients with early or newly diagnosed breast cancer with either node positive or negative status, no metastasis, no history of breast cancer surgery, and when imaging is done in the preoperative window. The reference standard had to be histopathology (sentinel lymph node biopsy and/or axillary dissection), and studies needed to report sufficient data to derive a 2 × 2 table of true positive (TP), false positive (FP), false negative (FN), true negative (TN), or equivalent accuracy metrics. We excluded studies restricted to recurrent or metastatic disease, studies enrolling male breast cancer patients, and reports without an appropriate histopathologic reference or without extractable accuracy data.

Search strategy

We searched databases of published studies PubMed/MEDLINE, Scopus, Web of Science, and Embase from inception to 31 December 2025, and screened references to select studies that met our inclusion criteria at first. Database-specific strategies comprised of vocabulary and keywords for breast cancer, axillary lymph nodes, ultrasound/sonography, MRI, and preoperative assessment. The full search strings and results by database are provided in Table S1.

Inclusion and exclusion criteria

Eligible studies were those that used histopathologic examination, either sentinel lymph node biopsy or axillary lymph node dissection as the reference standard, and reported sufficient data to calculate diagnostic performance measures such as true/false positives and negatives, sensitivity, specificity, or predictive values. Prospective and retrospective clinical studies, including cohort, case-control, and cross-sectional diagnostic accuracy designs, were considered. Only original human research published in English with full-text availability was included.

We included studies enrolling women with newly diagnosed primary invasive breast cancer undergoing preoperative axillary imaging (ultrasound and/or MRI) prior to definitive surgery, with histopathology (SLNB and/or ALND) as the reference standard. Early breast cancer was operationally defined as non-metastatic (Stage I-III) disease when staging information was reported. Studies were excluded if they focused on recurrent breast cancer, in situ-only carcinoma, or patients primarily treated with neoadjuvant systemic therapy unless diagnostic accuracy data were extractable for pre-treatment imaging.

Study selection process

Records were de-duplicated and screened in two stages (title/abstract, then full text) by two independent reviewers. Disagreements were resolved by discussion or, when consensus was not reached, by adjudication by a third reviewer. Reasons for full-text exclusion were documented. A PRISMA flow diagram of study selection is shown in Fig. 1.

Fig. 1.

Flow diagram indicating the selection of articles

Data extraction

Two reviewers independently extracted study characteristics (design, setting, sample size), participant details, index-test, and the reference standard. For each modality within a study, we captured or derived TP/FP/FN/TN. When multiple thresholds were reported, we used the authors’ primary definition. When contingency tables were not explicitly presented, counts were reconstructed from reported sensitivity/specificity and complementary metrics, then verified by a second reviewer. Discrepancies at the title/abstract and full-text stages were resolved by discussion; if consensus was not reached, a third senior reviewer adjudicated, with final decisions and exclusion reasons documented. Inter-reviewer agreement was assessed using Cohen’s kappa (κ) at both stages and interpreted using Landis-Koch benchmarks.

Risk of bias in studies (Quality assessment)

Risk of bias and applicability concerns were assessed using QUADAS-2 across four domains (patient selection, index test, reference standard, and flow/timing) and the first three domains for applicability [15]. Each domain was rated as low, high, or unclear risk, guided by QUADAS-2 signaling questions. Assessments were performed independently by two reviewers, with disagreements resolved by consensus. Results were summarized using traffic-light plot (Fig. 2).

Fig. 2.

Risk of Bias Assessment

Statistical analysis

Diagnostic accuracy for each imaging modality was summarized using 2 × 2 contingency tables (true positive [TP], false positive [FP], false negative [FN], true negative [TN]) against histopathology as the reference standard, as prespecified in the eligibility criteria. Where studies did not explicitly report a full 2 × 2 table, cell counts were algebraically reconstructed from the reported sample size and available accuracy indices (e.g., sensitivity, specificity, accuracy, PPV/NPV), and internal consistency was checked against the published metrics (allowing for rounding), with verification by a second reviewer. Reconstructed per-study counts are provided in Table S2.

For each study and modality, sensitivity (TP/[TP + FN]) and specificity (TN/[TN + FP]) were calculated or extracted from reported data, with 95% confidence intervals derived using the Wilson method. Primary meta-analytic pooling of sensitivity and specificity was performed using a generalized linear mixed-effects model (GLMM) for proportions on the logit scale, producing random-effects pooled estimates with 95% confidence intervals. From the pooled sensitivity and specificity, we derived positive and negative likelihood ratios (LR + and LR−) and the diagnostic odds ratio (DOR) for each modality. Between-study heterogeneity was quantified using I² (reported alongside pooled estimates).

To improve clinical interpretability, pooled sensitivity and specificity were translated into PPV and NPV at prespecified pre-test probabilities (10%, 20%, and 30%) using standard Bayesian relationships; PPV/NPV curves across prevalence are shown in Figure S1 (MRI) and Figure S2 (ultrasound). For direct (within-cohort) comparisons, studies reporting both MRI and ultrasound were analyzed using within-study log-odds ratios for sensitivity and specificity (MRI vs. ultrasound), which were pooled using random-effects meta-analysis and presented as pooled odds ratios (Table 4).

Table 4.

Head-to-head comparison (MRI vs US), pooled odds ratios

Outcome	k	Pooled OR (95CI)	I2	tau2	P overall
Sensitivity (MRI vs US)	12	0.96 (0.61-1.51)	0.8%	0.3555	0.842
Specificity (MRI vs US)	12	1.27 (0.74-2.18)	0.8%	0.3809	0.351

Random-effects meta-analysis of within-study log-odds ratios comparing MRI vs ultrasound. Outcomes shown for sensitivity and specificity

We assessed small-study effects/publication bias using Deeks’ funnel-plot asymmetry test, implemented as a regression of ln(DOR) on 1/√(effective sample size) weighted by effective sample size. Diagnostic trade-offs were additionally summarized using HSROC curves (Rutter-Gatsonis) from the bivariate random-effects model (Reitsma), with AUC reported (Fig. 5). Study-level sensitivity and specificity are displayed as modality-specific plots.

Fig. 6.

HSROC (Rutter-Gatsonis) from the bivariate random-effects model (Reitsma). On the left, MRI, group x studies: 26 - sensitivity = 57.5% - specificity = 90.9% - AUC = 0.863. On the right, US, studies: 17 - sensitivity = 59.4% - specificity = 88.9% - AUC = 0.843

Results

Characteristics of included studies

Twenty-nine studies, published between 2004 and 2025, were selected for this study (total sample size = 8544). Of the 29 included studies, 11 (37.9%) were prospective and 18 (62.1%) were retrospective. Regarding imaging modality, 5 studies (17.2%) used ultrasound alone, 11 studies (37.9%) used MRI alone, and 13 studies (44.8%) employed a combined MRI + ultrasound approach, indicating substantial variability in diagnostic strategies. The studies originated from countries across Asia, Europe, and North America, demonstrating broad international interest in improving non-invasive preoperative assessment of axillary lymph node status in early breast cancer (Table 1).

Table 1.

Summary of articles included in the present study

Study	Country	Design	Sample size	Modality	Age, Mean
Abe H (2013), [17]	USA	Retrospective	50	MRI, US	59.9
Ahn HS (2019), [18]	South Korea	Retrospective	74	MRI	51.5
Aktaş A (2022), [19]	Turkey	Retrospective	336	MRI, US	50.1
Arslan G (2016), [20]	Turkey	Prospective	35	MRI	52.86
Assing MA (2017), [16]	USA	Retrospective	271	MRI, US	58
Atallah D (2020), [21]	Lebanon	Retrospective	169	MRI	48
Barco I (2016), [22]	Spain	Prospective	1,505	MRI, US	58.5
Chayakulkheeree J (2019), [23]	Thailand	Retrospective	131	MRI	51.33
de Felice C (2014), [24]	Italy	Retrospective	26	MRI	56
Di Micco R (2025), [25]	Italy	Prospective	246	MRI	56.3
Ergul N (2015), [26]	Turkey	Prospective	24	MRI	47
Fernández A G (2011), [27]	Spain	Prospective	507	MRI, US	58
Hieken TJ (2013), [28]	USA	Retrospective	988	MRI, US	61.1
Jung NY (2015), [29]	South Korea	Retrospective	105	MRI	53.3
Kim SH (2017), [30]	South Korea	Retrospective	147	MRI	49.2
Motomura K (2015), [31]	Japan	Prospective	70	MRI	55
Panda SK (2019), [32]	India	Prospective	40	MRI, US	55
Riedel F (2021), [33]	Germany	Retrospective	611	US	NA
Rukanskiene D (2020), [34]	Lithuania	Prospective	227	US	59
Schumacher K (2021), [35]	USA	Retrospective	349	MRI, US	63
Stadnik TW (2006), [36]	Belgium	Retrospective	10	MRI	56
Upadhyaya VS (2019), [37]	Singapore	Retrospective	605	US	56.6
Valente SA (2012), [38]	USA	Retrospective	244	MRI, US	58
van Nijnatten TJA (2016)1, [39]	Netherlands	Retrospective	377	MRI, US	58
Wang J (2021), [40]	China	Prospective	85	US	42.36
Wang LJ (2025), [41]	Taiwan	Prospective	247	MRI, US	52.38
Xu Y (2018), [42]	China	Prospective	92	US	51
Yu Y (2021), [43]	China	Retrospective	803	MRI, US	48.7
Zhang H (2019), [44]	China	Retrospective	170	MRI, US	NA

¹Reported numbers for this study are calculated from the average of two MRI readers metrics

Risk of bias assessment

Risk of bias and applicability were evaluated using QUADAS-2 across patient selection, index test, reference standard, and flow/timing (Fig. 2). Overall study quality was mixed across the 29 included studies. Patient selection was the most frequent source of concern: only 4/29 studies were judged low risk, whereas 19/29 had some concerns and 6/29 were high risk, mainly because enrolment and exclusions were not fully described. Flow and timing showed the weakest profile (2/29 low risk; 17/29 some concerns; 10/29 high risk), suggesting that participant flow and the interval between imaging and histopathology were often insufficiently reported. By contrast, the index test and reference standard domains were more robust, with a larger proportion of studies rated low risk. Applicability concerns were generally low for the index test (22/29 low) and reference standard (23/29 low), while patient selection applicability was more variable (14/29 low).

Pooled performance

Both MRI and ultrasound showed moderate sensitivity and high specificity for preoperative axillary nodal assessment (Table 2; Figs. 3 and 4). In the primary random-effects GLMM, MRI achieved a pooled sensitivity of 61.1% (95% CI 48.3%−72.5%) and a pooled specificity of 92.0% (95% CI 88.3%−94.6%), corresponding to LR + 7.64, LR − 0.42, and a DOR of 18.1. Ultrasound yielded a pooled sensitivity of 59.8% (95% CI 49.0%−69.8%) and a pooled specificity of 90.3% (95% CI 84.6%−94.1%), with LR + 6.19, LR − 0.44, and a DOR of 13.9. Conventional inconsistency was minimal (I²=0.9% for both sensitivity and specificity). However, τ² (logit scale) and the 95% prediction intervals indicated that performance may vary across settings, particularly for sensitivity (MRI sensitivity PI 11.3%−95.1%, τ²=1.5769; MRI specificity PI 61.9%−98.8%, τ²=0.9525; ultrasound sensitivity PI 20.5%−89.6%, τ²=0.7504; ultrasound specificity PI 52.6%−98.7%, τ²=1.1070). Table 3

Table 2.

Study-level pooled diagnostic performance (sensitivity and specificity) for MRI and ultrasound (GLMM)

Modality	Metric	Pooled	95% CI	PI	I2	τ²	LR+	LR-	DOR
MRI	Sensitivity	61.1%	48.3% − 72.5%	11.3% −95.1%	0.9%	1.5769	7.64	0.42	18.1
MRI	Specificity	92.0%	88.3% − 94.6%	61.9% − 98.8%	0.9%	0.9525
US	Sensitivity	59.8%	49.0% − 69.8%	20.5% − 89.6%	0.9%	0.7504	6.19	0.44	13.9
US	Specificity	90.3%	84.6% − 94.1%	52.6% − 98.7%	0.9%	1.1070

Pooled sensitivity and specificity (95% CIs) estimated via a GLMM meta-analysis of proportions on the logit scale (primary approach). LR+, LR-,PI, and DOR are derived from the pooled Se/Sp

Fig. 3.

Study-level sensitivity and specificity (MRI)

Fig. 4.

Study-level sensitivity and specificity (Ultrasound)

Table 3.

PPV and NPV at prespecified pre-test probabilities

Prevalence	MRI PPV	MRI NPV	Ultrasound PPV	Ultrasound NPV
10%	45.9%	95.5%	40.7%	95.3%
20%	65.6%	90.4%	60.7%	90.0%
30%	76.6%	84.7%	72.6%	84.0%

PPV and NPV computed from pooled sensitivity and specificity (random-effects, logit scale). Values shown at 10%, 20%, and 30% pre-test probabilities

When translated to common clinical prevalence, MRI showed better PPVs compared to ultrasound at all given pre-test probabilities; on the other hand, NPVs were similar. For instance, at a 20% prevalence, MRI had a PPV of 65.6%, while ultrasound’s PPV was 60.7%. Regarding NPV at a 20% prevalence, the values were 90.4% and 90.0%, respectively. These patterns emphasize MRI’s superior rule-in performance without sacrificing the ability to rule out nodal disease in lower-risk settings (Table 4, Figures S1-S2).

Head-to-head comparison

Direct within-study contrasts reinforced these modality-level findings. Pooled odds ratios showed no difference in terms of sensitivity between MRI and ultrasound (OR 0.96; 95% CI, 0.61–1.51; p = 0.842; I²=0.8%). Regarding specificity, MRI had a better specificity, but the difference was non-significant (OR 1.27; 95% CI, 0.74–2.18; p = 0. 351; I²=0.8%). Thus, when both tests are applied to the same populations, MRI tends to be more specific, an attribute that reduces false positives and may spare unnecessary axillary interventions (Table 4, Figures S3-S4).

Publication bias

Potential reporting bias was assessed using Deeks’ funnel-plot asymmetry test. For MRI, the regression slope was statistically significant (p = 0.001), indicating asymmetry compatible with small-study effects. In contrast, the ultrasound funnel plot showed no evidence of asymmetry (p = 0.176). Because the test regresses ln(DOR) on the inverse square root of the effective sample size, asymmetry may reflect inflation of diagnostic odds ratios in smaller studies. This pattern suggests that pooled MRI estimates may be more vulnerable to selective publication or to systematic differences between smaller and larger cohorts. However, Deeks’ test is also influenced by some factors, including variation in MRI sequences, diagnostic thresholds for nodal positivity, and verification pathways. Accordingly, the finding should be interpreted as a cautionary signal rather than definitive proof of publication bias. Overall, the bias assessment supports greater confidence in the stability of the ultrasound summary estimates, while emphasizing the need to prioritize larger, methodologically well-described MRI studies when translating these results into clinical decision-making. Fig. 5

Fig. 5.

Deeks funnel (left: MRI, right: Ultrasound). MRI slope p-value = 0.001, Ultrasound slope p-value = 0.176. Non-significant slope suggests no small-study effects. Regression of ln (DOR) on 1/sqrt (ESS) weighted by ESS

Influence diagnostics were examined to identify studies with disproportionate impact on Deeks’ regression slope (Tables S3-S4). For MRI, the largest Cook’s distance was observed for Hieken et al. (2013), followed by Barco et al. (2016), indicating that a small number of heavily weighted cohorts contributed meaningfully to the observed funnel asymmetry. Several smaller MRI studies contained zero cells, but their Cook’s distances were low, suggesting limited leverage on the slope after continuity correction. For ultrasound, the most influential observations were again Hieken et al. (2013) and Barco et al. (2016), with moderate influence from Upadhyaya et al. (2019) and Schumacher et al. (2021). Overall, these diagnostics support interpreting Deeks’ findings as driven by a small subset of influential studies rather than uniformly across the evidence base.

Sensitivity analysis

A sensitivity analysis was performed by excluding the smallest 25% of studies based on effective sample size (Table 5). After exclusion, MRI estimates were derived from 19 studies (vs. 26 in the primary analysis) and showed a lower pooled sensitivity (52.6% vs. 61.1%), while specificity remained essentially unchanged (91.9% vs. 92.0%). In contrast, ultrasound results were stable: sensitivity was nearly identical after exclusion (59.6% vs. 59.8%), with a small reduction in specificity (88.8% vs. 90.3%). These findings suggest that MRI sensitivity may be more sensitive to small-study contributions, whereas ultrasound pooled performance is comparatively robust under this restriction.

Table 5.

Sensitivity analysis excluding the smallest 25% of studies (by ESS)

Modality	Studies: PrimaryAfter Exclusion	Sensitivity: Primary → After Exclusion	Specificity: Primary → After Exclusion
MRI	26 → 19	61.1% → 52.6%	92.0% → 91.9%
Ultrasound	17→ 13	59.8% → 59.6%	90.3% → 88.8%

Exclusion based on ESS derived from 2 × 2 tables (continuity correction applied if needed); pooled sensitivity and specificity re-estimated using the same GLMM approach

HSROC curves

HSROC analysis was used to summarize the joint performance of sensitivity and specificity across studies while accounting for threshold effects (Fig. 6). For MRI, the Rutter-Gatsonis HSROC derived from the bivariate random-effects model yielded a summary sensitivity of 57.5% and specificity of 90.9% across 26 studies, with an area under the curve (AUC) of 0.863. For ultrasound, the corresponding summary point was a sensitivity of 59.4% and specificity of 88.9% across 17 studies, with an AUC of 0.843. Both AUC values exceed 0.80, supporting good discriminative ability for preoperative nodal assessment. The modest separation between AUCs suggests slightly better overall discrimination for MRI, largely attributable to higher specificity rather than superior sensitivity. The HSROC curves also highlight an important trade-off: neither modality simultaneously achieves high sensitivity and high specificity, which aligns with the moderate pooled sensitivities reported in the primary synthesis. From a clinical perspective, this implies that imaging alone cannot replace histopathology for definitive axillary staging, particularly when missed metastasis would change management. Nevertheless, the high-specificity region of the MRI curve supports its role as a confirmatory test when a node is classified as suspicious, as it may reduce false-positive driven escalation of axillary surgery. Ultrasound, conversely, maintains slightly higher sensitivity and remains well suited for first-line screening and triage given its accessibility and real-time assessment. Taken together, the HSROC findings support a complementary pathway in which ultrasound is used broadly to identify patients who may require further assessment, and MRI is applied selectively to increase rule-in confidence and refine preoperative planning.

Discussion

In this diagnostic accuracy meta-analysis of 29 studies (n = 8,544), MRI and ultrasound showed broadly comparable performance for preoperative detection of axillary nodal metastasis. Sensitivity was moderate for both tests, whereas specificity was consistently high. Under the primary random-effects GLMM, pooled sensitivity was 61.1% for MRI and 59.8% for ultrasound, while pooled specificity was 92.0% and 90.3%, respectively, translating into stronger rule-in performance for MRI (LR + 7.64; DOR 18.1) than for ultrasound (LR + 6.19; DOR 13.9). Statistical inconsistency was low (I²=0.9%), but the estimated between-study variance (τ², logit scale) and wide 95% prediction intervals, most notably for sensitivity, suggest that real-world accuracy could differ meaningfully by institution and protocol. When mapped to clinically plausible pre-test probabilities (10%−30%), PPV increased as prevalence rose (MRI: 45.9% to 76.6%; ultrasound: 40.7% to 72.6), whereas NPV remained relatively stable and high across scenarios (Table 3). In direct head-to-head comparisons (k = 12), MRI and ultrasound did not differ in sensitivity (OR 0.96, 95% CI 0.61–1.51), and MRI showed a non-significant tendency toward higher specificity (OR 1.27, 95% CI 0.74–2.18).

The pooled sensitivity estimates in our study, 61.1% for MRI and 59.8% for US, are largely consistent with Literature. A systematic review by Alvarez et al. found MRI sensitivity ranging from 60 to 80%, mirroring our pooled MRI sensitivity of 61.1% [45].

Importantly, our study further clarifies the specificity advantage of MRI. Earlier meta-analyses have reported MRI specificities between 85% and 93%, but without a consistent head-to-head comparison with US [46, 47]. Our pooled MRI specificity of 92.0% not only confirms the strong rule-in value of MRI but also demonstrates a non-significant superiority compared to US (90.3%) in direct comparison. This aligns with the findings from Baxter et al., who showed MRI offered improved specificity when high-resolution sequences or diffusion-weighted imaging were implemented [48]. Our meta-analysis strengthens this observation through larger combined data and standardized statistical modeling.

The diagnostic odds ratio (DOR), a comprehensive measure integrating sensitivity and specificity, also favored MRI (18.1 vs. 13.9 for US), consistent with prior work. A meta-analysis by Li et al. found MRI DOR values around 20–25, and US around 10–15, closely matching our results and reinforcing the higher discriminatory performance of MRI [49]. These comparative findings support the use of MRI, particularly when higher diagnostic certainty is required, such as in cases with ambiguous US findings or in patients for whom accurate pre-operative mapping is critical.

A particularly valuable contribution of this meta-analysis is the predictive value modeling across clinically relevant pre-test probabilities. Prior studies rarely explored PPV/NPV across varying prevalence. Our analysis shows that MRI consistently produces higher PPVs at all examined nodal metastasis prevalence (10%, 20%, and 30%). For instance, at 30% prevalence, the PPV for MRI is 76.6%, compared to 72.6% for US. These differences have meaningful clinical implications: when MRI identifies suspicious nodes, the likelihood of true metastasis is substantially higher. Conversely, both modalities achieve similarly high NPVs, above 87% across all prevalence levels, indicating strong utility for ruling out nodal disease. This finding aligns with an earlier study by Kim et al., who highlighted the high negative predictive value of both modalities for excluding significant metastatic burden [50].

Because PPV and NPV are prevalence-dependent, we modeled predictive values across clinically plausible pre-test probabilities using Bayes’ rule, rather than reporting a single “fixed” PPV/NPV that may not generalize across settings [Bours, 2021 #56]. We selected 10%, 20%, and 30% to represent low-, intermediate-, and higher-risk pathways encountered in practice. For example, in patients with small tumors (≤ 2 cm) and negative preoperative axillary ultrasound, the SOUND trial still reported a 13.7% rate of nodal involvement, supporting a 10–15% scenario [Gentilini, 2023 #57]. Conversely, broader clinically node-negative populations often have nodal positivity closer to one quarter overall (i.e., ~ 74% sentinel-node negative), making 20–30% a reasonable bracket for symptomatic presentations or cohorts enriched with higher-risk features [Yu, 2022 #58]. In our study, this prevalence-based framing clarified the clinical trade-off: MRI’s higher specificity translated into consistently higher PPV than ultrasound across all modeled prevalences, while NPVs remained very similar; at 20% prevalence, PPV was 65.6% vs. 60.7% and NPV 90.4% vs. 90.0% (MRI vs. US).

The direct head-to-head comparison incorporated in our meta-analysis is in line with priori explained findings. While many prior meta-analyses evaluated MRI or US independently, relatively few studies directly contrasted both modalities in paired cohorts. Our inclusion of twelve such studies provides clearer insight: MRI and US do not differ significantly in sensitivity, but MRI is tend to be more specific. This nuanced comparative perspective was lacking in earlier reviews and is essential for evidence-based decision-making.

Integrating imaging performance with explicit clinical decision pathways is essential for translating diagnostic accuracy into safe axillary de-escalation. In our meta-analysis, ultrasound and MRI demonstrated only moderate sensitivity for detecting nodal metastasis (59.8% and 61.1%, respectively), whereas specificity was high, particularly for MRI (90.3% for ultrasound and 92.0% for MRI), supporting imaging as a pragmatic “rule-in” rather than a definitive “rule-out” strategy in isolation. This performance profile becomes more actionable when embedded in structured preoperative algorithms that define thresholds for escalation, confirmatory biopsy, or selective omission of surgical staging. For example, Chan and colleagues showed that adding MRI meaningfully altered management: MRI led to cancellation of breast-conserving surgery/IORT in a subset of patients, including cases where axillary nodal metastases were detected only on MRI, while also introducing measurable system burdens such as longer waiting time [51]. These decision-level effects underscore that the clinical value of imaging is not limited to summary accuracy metrics but also includes how it redirects treatment intensity. Importantly, Argentiero and co-authors emphasize that diagnostic strategies should not follow a single template and argue for integrating biological profiling into multistep work-ups, recognizing microenvironmental and molecular heterogeneity as key drivers of disease behavior [52]. Future research in axillary staging should therefore prioritize validated multimodal prediction frameworks that combine imaging with clinicopathologic and biologic variables to improve false-negative control while supporting individualized de-escalation [53].

Despite the strengths of our analysis, including advanced GLMM modeling, predictive value assessment, and publication bias evaluation, certain limitations must be acknowledged. Variations in imaging techniques, operator expertise, nodal size thresholds, and histopathological reference standards may introduce heterogeneity across studies. Some MRI studies used conventional sequences only, whereas others incorporated diffusion-weighted imaging or dedicated axillary protocols, contributing to variability. Additionally, MRI availability, higher cost, and longer acquisition times remain barriers in resource-limited settings. We couldn’t comprehensively evaluate the impact of imaging protocols heterogeneity because the key technical and interpretive details were not consistently reported and it is another main limitation. For MRI, reporting was similarly variable with respect to field strength, coil setup, axillary coverage, sequence selection (conventional sequences vs. diffusion-weighted imaging and/or dedicated axillary protocols), and reader blinding to histopathology. This incomplete and inconsistent reporting limited descriptive synthesis of protocol factors and precluded robust subgroup analyses or meta-regression to explain residual variability. For ultrasound, studies varied in criteria for nodal positivity (e.g., cortical thickness thresholds, hilar effacement, morphology, and Doppler features), operator experience, and whether suspicious nodes underwent confirmatory needle sampling.

Conclusion

MRI and ultrasound each play distinct roles in the pre-operative assessment of axillary lymph node status in early breast cancer. Ultrasound offers slightly higher sensitivity and can be effectively used as an initial screening tool. MRI demonstrates non-significant higher specificity, greater diagnostic odds ratio, and superior predictive value, making it a valuable confirmatory modality when precise nodal evaluation is required. The combined use of both imaging strategies may improve diagnostic accuracy, guide personalized surgical planning, and reduce unnecessary axillary procedures. Future standardized, prospective trials are warranted to refine decision-making algorithms and further optimize axillary staging.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

US

Ultrasound

MRI

Magnetic resonance imaging

ALN

Axillary lymph node

ALND

Axillary lymph node dissection

AUC

Area under the curve

CI

Confidence interval

DOR

Diagnostic odds ratio

ESS

Effective sample size

FN

False negative

FP

False positive

TN

True negative

TP

True positive

GLMM

Generalized linear mixed-effects model

HSROC

Hierarchical summary receiver operating characteristic

I²

I-squared heterogeneity statistic

IORT

Intraoperative radiotherapy

Κ

Cohen’s kappa coefficient

LR+

Positive likelihood ratio

LR-

Negative likelihood ratio

MEDLINE

Medical Literature Analysis and Retrieval System Online

PPV

Positive predictive value

NPV

Negative predictive value

OR

Odds ratio

PI

Prediction interval

PRISMA

Preferred Reporting Items for Systematic Reviews and Meta-Analyses

QUADAS-2

Quality Assessment of Diagnostic Accuracy Studies-2

SLNB

Sentinel lymph node biopsy

SOUND trial

Sentinel Node vs. Observation After Axillary UltraSound trial

τ²

Between-study variance (tau-squared)

Author contributions

Conception and Design: HJT, RB, FS, FNR, DY, SMAnalysis and interpretation of data: FNR, SS, ZA, DYData collection: RB, FS, FNR, SS, DYRisk of Bias Assessment: DY, FNR, HTJParticipating in drafting or revising the article: HJT, RB, FS, FNR, SS, DY, ZASupervision, Project Administration, Review & Editing: DY and ZA.

Funding

None.

Data availability

The dataset generated or analyzed in this study is available from the corresponding author upon a reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Consent for publication

Not applicable.

Clinical trial registration number

Not applicable.