Ultrasound-based deep-learning models for axillary nodal assessment show 0.89 specificity in breast cancerDeep learning models show promise for breast cancer ultrasound scans

Frontiers in Medicine Published October 23, 2026 DOI ↗ Editorial oversight: Dr. Julia Lee, PhD · Oncology, Genomics & Drug Development

AI-generated summary of the cited source, checked by automated accuracy review. How we work

Key Takeaway

Note that ultrasound-based DL models show 0.89 specificity but should be used as adjuncts to surgical pathology.

This meta-analysis evaluated the diagnostic performance of deep-learning (DL) models using ultrasound as the sole imaging modality for pretreatment axillary or sentinel nodal status in patients with breast cancer. The analysis pooled data from 11 studies involving 1,027 validation patients to compare DL models against the reference standard of surgical pathology.

The meta-analysis reported a sensitivity of 0.75 (95% CI 0.66 to 0.81) and a specificity of 0.89 (95% CI 0.83 to 0.93). Other reported metrics included a positive likelihood ratio of 6.60 (95% CI 4.29 to 10.42), a negative likelihood ratio of 0.29 (0.21 to 0.38), and a diagnostic odds ratio of 23.12 (12.89 to 41.50).

Several limitations were noted, including wide prediction intervals, heterogeneous anatomical inputs and preprocessing, and incompletely reported definitions of nodal positivity. Additionally, the studies were geographically concentrated. These findings suggest that while ultrasound-based DL models show moderate sensitivity and high specificity, they should be viewed as adjuncts to, rather than replacements for, pathological axillary staging.

How this fits prior evidence

This meta-analysis addresses a gap in the technological assessment of preoperative axillary nodal staging in breast cancer. While previous coverage has focused on systemic treatments like bevacizumab-containing regimens and molecular pathways such as JAK-STAT or miR-155, this study specifically evaluates the diagnostic performance of deep-learning algorithms in ultrasound imaging.

When doctors treat breast cancer, they must determine if the cancer has spread to the lymph nodes under the arm, known as the axillary region. This step is vital for planning the right treatment. A large review of 11 different studies involving over 1,000 patients looked at how well deep-learning models—a type of advanced computer software—could identify these nodes using only ultrasound images.

The analysis found that these computer models had a high specificity of 0.89, meaning they were very good at correctly identifying patients who did not have positive nodes. However, the sensitivity was lower at 0.75, meaning they missed some cases. Because of these mixed results, experts suggest these tools should be used to help doctors, not replace the standard surgical check of the tissue.

It is important to note that the data comes from a variety of different settings and study types. Because the studies used different ways to define and process images, the results are not yet perfect for every hospital. These tools are currently best viewed as a helpful extra layer of information for doctors during the staging process.

What this means for you:

AI models show high accuracy in identifying clear lymph nodes but should assist, not replace, standard testing.

Common questions

How accurate are these AI models at detecting cancer?

The models showed a sensitivity of 0.75 and a specificity of 0.89. This means the models were very good at identifying patients without positive nodes, but they were less consistent at catching every case of positive nodes. Because of this, they are currently recommended as a tool to help doctors rather than a replacement for standard testing.

What is the role of ultrasound in this process?

In this study, ultrasound was used as the only imaging method for the deep-learning models to determine the status of lymph nodes. The goal is to see if these computer models can help doctors assess the axillary area before surgery to help plan the best treatment for breast cancer patients.

Are there any risks or limitations to using these models?

The study noted several limitations, including a wide range of prediction intervals and different ways that different studies defined what a positive node was. Because of these variations and the different ways images were processed, these models are currently intended to be used alongside standard surgical pathology, not as a standalone tool.

Study Details

Study typeMeta analysis

EvidenceLevel 1

PublishedOct 2026

View Original Abstract ↓

BackgroundDeep-learning (DL) models may extract nodal-risk information from breast or axillary ultrasound, but the relevant studies differ in anatomical input, preprocessing, reference standard, and validation design.MethodsWe updated PubMed, Web of Science Core Collection, and Cochrane CENTRAL, with supplementary OpenAlex and Crossref searches and repeated backward and forward citation searching through 24 August 2026. Eligible studies evaluated a deep-learning model using ultrasound as the sole imaging modality for pretreatment binary axillary or sentinel nodal status in breast cancer, used surgical pathology as the reference standard, and reported an independent patient-level validation cohort with an exact or uniquely reconstructable 2 × 2 table. One cohort per study was selected using a prespecified hierarchy. Sensitivity and specificity were synthesized with a bivariate random-effects model and an HSROC curve.ResultsEleven studies comprising 1,027 validation patients were included. Pooled sensitivity was 0.75 (95% CI 0.66–0.81; 95% prediction interval 0.50–0.90) and pooled specificity was 0.89 (95% CI 0.83–0.93; prediction interval 0.66–0.97). The summary HSROC area was 0.81. The positive likelihood ratio was 6.60 (95% CI 4.29–10.42), the negative likelihood ratio was 0.29 (0.21–0.38), and the diagnostic odds ratio was 23.12 (12.89–41.50). Descriptive univariate I² values were 64.2% for sensitivity and 73.8% for specificity. Deeks’ test did not indicate funnel-plot asymmetry (P = 0.677).ConclusionsUltrasound-based DL models showed moderate sensitivity and high specificity in selected validation cohorts. Wide prediction intervals, heterogeneous anatomical inputs and preprocessing, incompletely reported definitions of nodal positivity, and the geographic concentration of studies limit immediate clinical transportability. These models should be evaluated as adjuncts to, rather than replacements for, pathological axillary staging.Systematic review registrationhttps://www.crd.york.ac.uk/PROSPERO/, identifier CRD42024589342.

Ultrasound-based deep-learning models for axillary nodal assessment show 0.89 specificity in breast cancerDeep learning models show promise for breast cancer ultrasound scans

How this fits prior evidence

Common questions

Study Details

First-line immune checkpoint inhibitor combinations improve overall survival in extensive-stage small-cell lung cancer

Immune checkpoint inhibitors combined with chemotherapy improve small cell lung cancer survival

Clinical research that matters. Delivered to your inbox.

Ultrasound-based deep-learning models for axillary nodal assessment show 0.89 specificity in breast cancerDeep learning models show promise for breast cancer ultrasound scans

How this fits prior evidence

Common questions

More on Breast Cancer

Study Details

First-line immune checkpoint inhibitor combinations improve overall survival in extensive-stage small-cell lung cancer

Immune checkpoint inhibitors combined with chemotherapy improve small cell lung cancer survival

Clinical research that matters. Delivered to your inbox.

Related in Oncology

From Other Specialties