AI Reads Ultrasound to Predict Breast Cancer Spread Before Surgery

Deep learning models that read ultrasound images can flag breast cancer spread to the armpit lymph nodes with moderate sensitivity and high specificity, according to a systematic review and meta-analysis published in Frontiers in Oncology. Pooling data from 11 independently evaluated validation cohorts totaling 1,027 patients, researchers led by Tianxin Liu of the First Affiliated Hospital of Shihezi University calculated a pooled sensitivity of 0.75 (95% confidence interval 0.66 to 0.81) and a pooled specificity of 0.89 (95% confidence interval 0.83 to 0.93). In plain terms, the algorithms caught roughly three out of four patients whose lymph nodes harbored metastatic disease while correctly clearing about nine out of ten patients whose nodes were clean. The summary area under the hierarchical summary receiver operating characteristic curve, a single number that condenses the trade-off between sensitivity and specificity across all thresholds, came out at 0.81, a level the authors characterize as promising but not yet decisive.

The stakes behind those percentages are considerable. Axillary nodal status remains central to breast cancer staging, prognosis, and locoregional treatment planning, and the field has been moving steadily toward selective de-escalation of axillary surgery in defined early-stage settings. The current reference standard is surgical pathology, obtained either through sentinel lymph node biopsy or axillary lymph node dissection. Both procedures are invasive, and sentinel biopsy samples only a selected subset of nodes. A reliable non-invasive triage tool applied before surgery could refine preoperative risk assessment while preserving pathology for the definitive decisions that matter most. That is the clinical gap these artificial intelligence models are trying to fill, and the reason the review team insisted on surgical pathology as the benchmark against which every model was judged.

Conventional axillary ultrasound, read by a radiologist, has long offered a specific but insensitive first look. In a prior bivariate meta-analysis of pretherapeutic imaging cited by the authors, visual ultrasound assessment achieved a pooled sensitivity of just 0.51 (95% confidence interval 0.43 to 0.59) with specificity of 1.00 (0.99 to 1.00), meaning it rarely produced false alarms but missed a clinically important proportion of nodal disease. The deep learning models reviewed here work differently and, crucially, ask a different question. Most of them analyze images of the primary breast tumor itself and infer the probability that cancer cells have already migrated to the nodes, rather than directly inspecting the nodes. The authors are explicit that these differences in anatomical target, patient spectrum, threshold, and reference standard preclude any claim of superiority over conventional ultrasound without within-patient head-to-head studies.

The review itself was built with unusual methodological discipline. Registered in PROSPERO under identifier CRD42024589342 and reported according to PRISMA 2020 and PRISMA-DTA standards, it set a narrow diagnostic question: among patients with untreated breast cancer, how accurately do deep learning models using ultrasound as the sole imaging modality classify binary axillary or sentinel nodal status in an independent patient-level validation cohort against surgical pathology? To avoid optimistic multiplicity, the team selected only one cohort and one model per study, following a prespecified hierarchy that favored external validation, then prospective or temporal validation, then an independent internal holdout. They deliberately did not choose the threshold or model with the highest Youden index, a choice that guards against inflated performance estimates. Every reconstructed two-by-two table was checked by recomputing sensitivity, specificity, cohort size, and disease prevalence.

The search machinery behind the analysis was extensive. The completed searches identified 531 source records across PubMed, Web of Science Core Collection, Cochrane CENTRAL, and supplementary OpenAlex and Crossref queries. After removal of 295 duplicates, 236 records were screened, 200 were excluded at title and abstract stage, and 36 reports were assessed in full text, of which 25 were excluded. Eleven studies published between 2020 and 2026 made the final cut. All of them were conducted in China, a geographic concentration the authors flag as a major limitation for transportability. Eight of the selected cohorts were external, prospective, or temporal validations, while three were independent internal holdouts. Eight models used only primary-tumor ultrasound, three incorporated axillary-node images, ten relied on static B-mode input, and one processed ultrasound video.

The diagnostic arithmetic is where the review delivers its most quotable numbers. The positive likelihood ratio was 6.60 (95% confidence interval 4.29 to 10.42), the negative likelihood ratio was 0.29 (0.21 to 0.38), and the diagnostic odds ratio was 23.12 (12.89 to 41.50). Translated into clinical terms using Fagan post-test probability calculations, at a pretest probability of nodal involvement of 20 percent, a positive model result raised the post-test probability to 62.3 percent, while a negative result lowered it to 6.7 percent. At a pretest probability of 30 percent, the corresponding figures were 73.9 percent after a positive result and 11.0 percent after a negative one. The authors interpret this asymmetry carefully: a positive result may provide useful rule-in information, but the negative likelihood ratio is not low enough to justify replacing pathological staging when a missed nodal metastasis would change management.

Beneath the pooled averages lies substantial dispersion. Descriptive univariate heterogeneity was I-squared of 64.2 percent for sensitivity and 73.8 percent for specificity, and the random-effects standard deviations in the validation fit were 0.51 for logit sensitivity and 0.68 for logit specificity, with an estimated correlation of minus 0.26. More striking are the 95 percent prediction intervals, which describe what a future study in a new setting might plausibly achieve: 0.50 to 0.90 for sensitivity and 0.66 to 0.97 for specificity. Those ranges are materially wider than the confidence intervals around the means, signaling that performance in an unseen center could differ substantially from the pooled estimate. Exploratory subgroup analyses, based on small numbers of studies and without formal interaction testing, hinted that pure deep learning models were more specific but less sensitive (sensitivity 0.65, specificity 0.94) than hybrid models combining deep features with radiomics or clinical variables (sensitivity 0.79, specificity 0.85).

Risk of bias assessment with QUADAS-2 revealed a mixed picture. Reference-standard and flow-and-timing domains were generally low risk because surgical pathology was used and validation data were linked at the patient level. Patient selection was frequently unclear in retrospective cohorts, however, and two studies raised applicability concerns: one enrolled only patients aged 75 or older, and another selected patients with suspicious axillary ultrasound findings. Index-test risk was usually unclear because prespecification of classification thresholds was incompletely described, and one study was judged high risk for threshold-related concerns. A further wrinkle is the definition of nodal positivity itself. Explicit handling of isolated tumor cells and micrometastases was often missing across the included reports, meaning the positive class may not have been identical from cohort to cohort, even though restricting a subgroup to studies explicitly using sentinel biopsy or dissection produced similar mean estimates.

The limitations extend beyond study design. Embase could not be updated because institutional access was unavailable, and although OpenAlex and Crossref were searched as supplementary public sources, the authors note these do not replace Embase. Sparse subgroups precluded reliable meta-regression, and Deeks’ funnel-plot asymmetry test, which showed no evidence of small-study effects (P = 0.677), had limited power with only 11 studies, so absence of statistical evidence should not be read as proof that reporting bias is absent. The analysis estimates diagnostic accuracy, not patient outcomes or the safety of omitting surgery, an important boundary on what these numbers can promise.

Where does this leave the technology? The authors conclude that ultrasound-based deep learning models should be evaluated as adjuncts to, rather than replacements for, pathological axillary staging. The pooled figures support potential use in triage or as a second reader, but the appropriate action threshold depends on whether the intended pathway is selecting patients for needle sampling, for sentinel biopsy, or for observation. The research agenda they lay out is demanding: prospective multicenter impact studies with the model and threshold locked before external testing, direct comparison against radiologist-performed axillary ultrasound under the same reference standard, standardized image preprocessing in line with initiatives such as the Image Biomarker Standardization Initiative, and complete reporting under medical imaging AI guidance such as CLAIM. Until those studies arrive, the algorithms remain promising assistants in the operating theater’s waiting room, not gatekeepers of the scalpel.

Subject of Research: Diagnostic accuracy of deep learning models using ultrasound for preoperative axillary lymph node status in breast cancer

Article Title: Diagnostic performance of deep learning models using ultrasound for preoperative axillary nodal status in breast cancer: a systematic review and meta-analysis

Article References: Liu, T., Yuan, X., Wang, J., He, Y., Liu, Y., Yan, X., Cao, C., Chen, M., Li, W., Li, J., & Cui, X. (2026). Diagnostic performance of deep learning models using ultrasound for preoperative axillary nodal status in breast cancer: a systematic review and meta-analysis. Frontiers in Oncology, 16, Article 1925423. https://doi.org/10.3389/fonc.2026.1925423

Image Credits: AI Generated

DOI: 10.3389/fonc.2026.1925423

Keywords: breast cancer, deep learning, ultrasound, axillary lymph nodes, sentinel lymph node biopsy, diagnostic accuracy, meta-analysis, artificial intelligence, radiomics, cancer staging, medical imaging AI, Diagnostic