This study was performed in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 guidelines (Checklist 1) and was registered in the PROSPERO (Prospective Register of Systematic Reviews) database (ID CRD42024609828).

The eligibility criteria were established as the basis for subsequent literature screening as shown in Table 1.

Web of Science, Cochrane Library, PubMed, and Embase were searched from their inception up to January 26, 2026. The search strategy was designed by combining free-text terms and subject headings. There were no temporal or geographical restrictions. The search strategy is illustrated in Table S1 in Multimedia Appendix 1.

Initially retrieved studies were imported into EndNote. After removing duplicates, the titles and abstracts of the remaining studies were further reviewed. After a full-text review, eligible studies were selected. Data extraction was performed using a standardized form. The extracted information included digital object identifier, publication year, patient source, country, title, study design, first author, number of ALNM cases, total cases, number of ALNM cases within the training cohort, total cases within the training cohort, number of ALNM cases within the validation cohort, total cases within the validation cohort, methods used to generate the validation cohort, whether region of interest segmentation was performed, model type, and whether a clinical comparison was conducted. Two reviewers independently selected the studies and extracted the data. Disagreements were resolved through consultation with a third reviewer.

The risk of bias (RoB) of the included studies was assessed using the Quality Assessment of Diagnostic Accuracy Studies [17]. This tool assessed overall bias and clinical applicability in 4 domains: index test, flow and timing, reference standard, and patient selection. Each domain was rated as low, high, or unclear RoB according to specific criteria. Two reviewers independently assessed the RoB and cross-checked the evaluation results. Disagreements were resolved through consultation with a third reviewer.

Data analysis was performed using Stata 15.0 (StataCorp LLC). A bivariate mixed effects (BME) model was used to explore the nonlinear relationship between sensitivity and specificity. The model estimated the following along with their corresponding 95% CIs: pooled sensitivity, specificity, area under the summary receiver operating characteristic curve (AUC), negative likelihood ratio (LR−), positive likelihood ratio (LR+), and diagnostic odds ratio (DOR). Publication bias was evaluated using the Deeks funnel plot. Sensitivity and specificity were derived from a 2×2 contingency table. Since most studies did not report contingency tables, the pooled sensitivity, specificity, precision, and case numbers were used for calculations. In the validation phase, a meta-analysis was conducted. When multiple validation sets were available within a single investigation, all were incorporated into the analysis. Subgroup analyses were performed, stratified by image type and validation set generation method. Furthermore, the analysis aggregated the outcomes from both internal and external validations for DL models developed using ultrasound, magnetic resonance imaging (MRI), and computed tomography (CT). Moreover, for studies with multiple validation cohorts, only the cohort with the highest Youden index was retained to ensure that a single validation cohort was retained in each study for the sensitivity analysis. The meta-analysis was then repeated with this reduced dataset. A P value <.05 was considered statistically significant.

A total of 2225 studies were selected from the databases, among which 521 duplicates were removed. Following a title and abstract screening, 1659 studies were further excluded. The full texts of the remaining 45 studies were assessed, and 17 studies were removed for the following reasons: inaccessible conference abstracts (n=4), absence of outcome indicators for assessing the accuracy of DL (n=6), sole focus on image segmentation (n=2), use of positron emission tomography-CT as the diagnostic gold standard (n=1), and lack of clear differentiation between sentinel and axillary lymph nodes (n=4). Ultimately, 28 studies [14,16,18-43] were included in the analysis. The study screening process is illustrated in Figure 1.

A total of 28 studies published between 2018 and 2026 were included. These studies involved 20,811 patients with BC, 7123 of whom had ALNM. Twenty-three of the studies were from China, 3 from the United States, 1 from Italy, and 1 from Korea. All of the studies used a case-control design. Seventeen of the studies were single-center, and 11 were multicenter. Regarding imaging modalities, 10 studies utilized MRI, 14 used ultrasound images, and 4 used CT scans. All 28 studies clearly described their validation methods. Twenty studies employed random sampling, 10 used external validation, and 3 utilized cross-validation. Twenty-three studies conducted manual segmentation, and 7 reported the comparisons between DL models and clinical physicians (Table 2).

Fifteen studies, comprising 34 diagnostic 4-fold tables, validated DL models constructed using Chinese populations, with an ALNM prevalence of 40%. The pooled results from the BME model were as follows: a sensitivity of 0.80 (95% CI 0.76‐0.84), a specificity of 0.85 (95% CI 0.80‐0.88), an LR+ of 5.3 (95% CI 4.1‐6.8), an LR− of 0.23 (95% CI 0.19‐0.28), a DOR of 23 (95% CI 16‐31), and an AUC of 0.89 (95% CI 0.86‐0.92; Figures S37 and S38 in Multimedia Appendix 1). The Deeks funnel plot revealed no significant publication bias (P=.08; Figure S39 in Multimedia Appendix 1). Assuming a prior probability of 0.4, a positive test result corresponded to a TP probability of 0.78 (Figure S40 in Multimedia Appendix 1).

Six studies, comprising 5 diagnostic 4-fold tables, validated DL models constructed using non-Chinese populations, with an ALNM prevalence of 38%. The pooled results from the BME model were as follows: a sensitivity of 0.79 (95% CI 0.59‐0.90), a specificity of 0.84 (95% CI 0.78‐0.89), an LR+ of 5.0 (95% CI 3.5‐7.3), an LR− of 0.25 (95% CI 0.12‐0.53), a DOR of 20 (95% CI 7‐53), and the AUC of 0.85 (95% CI 0.82‐0.88; Figures S41 and S42 in Multimedia Appendix 1). The Deeks funnel plot demonstrated no significant publication bias (P=.67; Figure S43 in Multimedia Appendix 1). Assuming a prior probability of 0.4, a positive test result corresponded to a TP probability of 0.77 (Figure S44 in Multimedia Appendix 1).

A sensitivity analysis was conducted by retaining only 1 validation cohort per study. For studies with multiple cohorts, the validation cohort with the highest Youden index was selected.

This meta-analysis revealed that DL approaches for detecting ALNM in BC primarily relied on MRI, ultrasound, and CT. The sensitivity and specificity for ultrasound-based DL models were 0.79 and 0.86, respectively. Meanwhile, the sensitivity and specificity of MRI-based models were 0.78 and 0.82, respectively. Furthermore, CT-based models demonstrated a sensitivity of 0.90 and a specificity of 0.88. The sensitivity of CT was superior to that of MRI and ultrasound, while no substantial differences in specificity were observed among the 3 modalities. These robust performance metrics validated the potential of DL as a complementary tool to conventional imaging assessments.

Currently, ML has gained widespread attention in the diagnosis and treatment of BC. Previous systematic reviews have evaluated the accuracy of ML models based on various imaging modalities. For instance, a meta-analysis by Jing Zhang et al [44], which included 13 studies involving 1618 patients, demonstrated that dynamic contrast-enhanced MRI radiomics had promising diagnostic performance in detecting ALNM and sentinel lymph node metastasis in patients with BC. However, only 2 of these studies employed DL methods, and the pooled sensitivity and specificity were 0.84 (95% CI 0.53‐0.96) and 0.65 (95% CI 0.31‐0.89), respectively. The limited number of studies constrained the interpretation of the results. Another meta-analysis by Chen et al [45], which included 14 studies, focused on ML-based MRI for diagnosing ALNM in patients with BC, and the pooled sensitivity and specificity were 0.79 (95% CI 0.74‐0.84) and 0.77 (95% CI 0.73‐0.81), respectively. This study primarily summarized models, such as support vector machines, logistic regression, and linear discriminant analysis, without specifically addressing the diagnostic accuracy of DL techniques. Eldaly et al [46] systematically reviewed radiomics approaches for detecting ALNM. Their results revealed that the AUC of the included studies varied from 0.72 to 0.93. However, they did not provide a detailed summary of diagnostic performance, thus limiting the assessment of its true accuracy. Liu et al [47] investigated the feasibility of AI algorithms based on CT and MRI for detecting ALNM in BC. They found that MRI-based methods demonstrated a sensitivity of 0.85 (95% CI 0.79‐0.90) and specificity of 0.81 (95% CI 0.66‐0.83), while CT-based methods showed a sensitivity of 0.88 (95% CI: 0.79‐0.94) and specificity of 0.80 (95%CI: 0.69‐0.88). A radiomics study [48] included MRI, ultrasound, CT, and X-ray mammography. A meta-analysis of 30 studies involving a total of 5611 patients was conducted. The pooled sensitivity and specificity of radiomics for detecting ALNM were 0.86 (95% CI 0.82‐0.88) and 0.79 (95% CI 0.73‐0.84), respectively, demonstrating strong overall diagnostic accuracy. Gong et al [48] comprehensively evaluated the performance of radiomics in detecting ALNM. Their results demonstrated an overall diagnostic accuracy of 23 (95% CI 16‐33) and a sensitivity of 0.86 (95% CI 0.82‐0.88). Radiomics serves as a promising noninvasive approach that can contribute to offering new quantitative modalities for disease diagnosis. However, their study did not strictly differentiate between types of ML, and different ML methods demonstrated varying performance in detecting positive events. Furthermore, there was no strict distinction between datasets in their study, with notable variations observed between the results in the training and validation sets.

With the advancement of AI, there has been a shift from traditional ML to DL. DL can analyze complex and diverse data, enabling personalized diagnosis and treatment that aligns with the current trend of precision medicine. Previous meta-analyses have lacked comprehensive evidence on DL models for detecting ALNM in patients with BC across various imaging modalities.

The present meta-analysis only included DL-based studies to minimize the variability of outcomes caused by different modalities. DL offers significant advantages over traditional ML with more advanced techniques for image processing, contributing to the development of diagnostic models for diseases [37]. Our findings suggested that image-based DL for detecting ALNM in BC demonstrated favorable diagnostic performance. MRI, ultrasound, and CT are commonly used techniques for detecting ALNM in BC. However, MRI is not necessary or routinely used in BC diagnosis due to its higher cost. Our findings reveal that CT-based models achieved higher sensitivity, specificity, and AUC than models based on ultrasound and MRI. However, since only 4 studies on CT were included, the results should be validated using larger datasets. Therefore, future research should develop more CT-based diagnostic models using AI to expand its role in diagnosing BC.

The trade-off between sensitivity and specificity for ultrasound and MRI highlights their complementary roles in clinical practice. The high sensitivity of ultrasound models effectively reduces false negatives by excluding metastasis, whereas the high specificity of MRI models improves confirmation of positive cases by minimizing false positives. These findings support the development of an integrated multimodal DL strategy, with ultrasound as an initial screening tool and MRI for confirming difficult cases. However, real-world validation is required.

Before developing DL-based intelligent diagnostic tools, several challenges require attention. First, the impact of different imaging protocols and preprocessing methods on DL performance must be thoroughly examined. Second, DL models typically require large image datasets. Therefore, future studies should incorporate more extensive imaging data to ensure the robustness of the DL model. Third, the impact of population heterogeneity on model performance should be carefully considered. Future research should include multicenter, multiethnic cohorts when developing DL models. Fourth, external validation is essential to enhance the credibility and generalizability of these DL models. Nevertheless, most included studies relied solely on internal validation, making it difficult to conduct external validation across multiple centers. Among the 28 studies, the majority employed single-center internal validation. Therefore, future research should incorporate more multicenter studies to confirm the real-world applicability of these models.

This is the first study to systematically investigate the feasibility of DL for detecting ALNM in BC. However, several limitations should be noted. First, despite a comprehensive literature search, the number of included studies was limited, restricting in-depth analysis and the generalizability of the findings. Second, most included studies relied on random sampling validation and lacked diverse external validation, undermining the overall reliability and generalizability of the results. Third, only a few studies directly compared DL models with clinical experts, and the available data were too limited to support a robust effect size analysis. As a result, it remains uncertain whether DL surpasses clinical experts in diagnostic performance. Fourth, many studies did not provide detailed descriptions of critical aspects, such as image acquisition protocols, segmentation techniques, network architectures and inputs, and reference standards (including biopsy methods and reference intervals). Fifth, this meta-analysis relied solely on data reported in the published articles and their supplementary materials. We did not contact the corresponding authors regarding missing or incomplete data. This may have resulted in the exclusion of some potentially eligible studies, which could introduce bias if the missing data differ systematically from the included data. Sixth, most included studies were based on retrospective evidence. Therefore, future research should focus on standardizing image processing methods and conducting prospective, multicenter external validations to develop DL models with enhanced robustness and broader applicability. Finally, this meta-analysis incorporated validation cohorts from the included studies. For studies that provided multiple validation cohorts, each cohort was included as an independent data point in the analysis. This approach may violate the inherent assumption of independence in the standard BME model. While we conducted a sensitivity analysis by extracting only a single validation cohort per original study, this alternative method may also introduce a certain degree of publication bias. Consequently, although DL models show promise for detecting ALNM, the current findings should be interpreted with caution. Future prospective studies with preregistered protocols are needed to verify these results and minimize publication bias.

This study revealed that DL models based on medical imaging showed promising accuracy in detecting ALNM in patients with BC. These findings may provide an evidence-based foundation for developing intelligent diagnostic tools. However, the findings were primarily based on models developed from single-center, retrospective datasets. Therefore, future studies should adopt a series of strategies, including but not limited to expanding the sample size, integrating multimodal imaging for joint modeling, and conducting prospective validation, to further enhance the performance of DL models while ensuring their applicability and clinical utility.