AI Model Predicts Breast Cancer Spread to Lymph Nodes Before Surgery

For women diagnosed with invasive breast cancer, one of the most consequential questions asked in the first days after diagnosis is deceptively simple: has the disease reached the lymph nodes under the arm? The answer shapes surgical planning, staging and the intensity of subsequent treatment, yet today it can only be answered definitively by surgically removing sentinel lymph nodes or performing a full axillary dissection. A new study published in BMC Medical Imaging by He Nie, Huarui Tan, Zhen Cao and Xiaozheng Cao explores whether a carefully engineered artificial-intelligence pipeline can answer that question before a single incision is made, using only information that is already available at the time of staging.

The research team assembled five complementary streams of data from sixty-three women with histologically confirmed invasive breast carcinoma who underwent baseline fluorine-18 fluorodeoxyglucose positron-emission tomography combined with computed tomography, the scan commonly abbreviated as 18F-FDG PET/CT. Forty-two of the patients were node-positive and twenty-one were node-negative, and forty of the sixty-three, roughly sixty-three percent, subsequently received neoadjuvant systemic therapy. From the tumour volume delineated on co-registered PET and CT images, the investigators extracted 2,167 radiomics features compliant with the Image Biomarker Standardisation Initiative, 12,288 deep-learning features obtained through transfer learning from a ResNet-50 convolutional neural network, eleven volumetric metabolic parameters, seven routine clinicopathological variables and seven blood-count-derived indices of systemic inflammation and nutritional status.

What distinguishes this study from many similar multimodal efforts is not the ambition of its feature set but the rigour of its evaluation. In machine-learning medicine, a notorious pitfall is information leakage: when preprocessing steps such as feature selection are fitted on the entire dataset before validation, the model inadvertently peeks at the very cases it is supposed to predict, inflating apparent performance. The authors designed their pipeline to prevent this. Median imputation of missing values, variance filtering, standardisation, univariate screening by area under the receiver-operating-characteristic curve, collinearity pruning and least absolute shrinkage and selection operator, or LASSO, feature selection were all fitted exclusively within the training folds of repeated stratified five-fold cross-validation, repeated ten times.

The modelling search itself was exhaustive. Five single-block models, one for each data stream, were compared against three fusion strategies: stacking, feature concatenation and a parameter-free probability mean. Four base learners were tested, and critically, the complete choice of data block, fusion strategy, learning algorithm and decision threshold was repeated inside a nested cross-validation scheme. Uncertainty was quantified by patient-level bootstrap, models were compared using paired bootstrap tests, and the authors additionally assessed calibration, decision-curve net benefit, the incremental value of imaging over clinicopathological variables, the handling of missing data and robustness to contour perturbation of the tumour outlines.

The headline result is sobering but honest. The nested cross-validated area under the curve of the entire modelling procedure was 0.711, with a ninety-five percent confidence interval of 0.609 to 0.805. The probability-mean fusion, the strategy most frequently selected in the inner loops of the nested validation, reached 0.795 on the same splits, with a confidence interval of 0.678 to 0.891. The gap between these two figures, approximately 0.08, is a transparent quantification of model-selection optimism, the amount by which the best-performing configuration overstates true performance simply because it was chosen from many candidates. At thresholds set on training folds, the fused model achieved a sensitivity of 0.80 and a specificity of 0.60, and its calibration was imperfect, with a slope of 1.53, meaning predicted probabilities did not align perfectly with observed frequencies.

Perhaps the most clinically provocative finding concerns the value of imaging itself. The clinicopathological variables alone achieved a cross-validated AUC of 0.681. Adding the PET/CT-derived information increased the AUC by 0.107, but the confidence interval for that increment ranged from minus 0.023 to 0.247, and the paired comparison yielded a p-value of 0.12. In plain terms, the study could not statistically demonstrate that the expensive imaging features added incremental predictive value over routinely collected clinical data in this cohort. Decision-curve analysis reinforced the caution: the fused model offered a net benefit over a treat-all strategy only at threshold probabilities of about 0.4 or higher, a relatively narrow window of clinical usefulness.

The single-block results add further texture. Deep-learning features extracted by the ResNet-50 network were the strongest individual predictor, reaching an AUC of 0.750, while the haematological inflammation-nutrition indices performed worst at 0.590, a confidence interval that included 0.5, the level of a coin flip. This hierarchy is biologically plausible, since deep convolutional networks can capture texture and morphological patterns in tumour glucose metabolism that correlate with aggressive biology and nodal spread, whereas systemic inflammation markers reflect host factors that are only loosely coupled to lymphatic dissemination.

The authors also stress-tested their features in ways that many radiomics studies neglect. Under a stringent one-voxel perturbation of the tumour contours, only twenty-three percent of the radiomics features showed an intraclass correlation coefficient of at least 0.75, a standard benchmark for reproducibility. Importantly, restricting the radiomics block to these robust features did not reduce model performance, suggesting that the pipeline can be pruned to stable features without loss. Discrimination was notably poor in the subgroup of patients who received neoadjuvant systemic therapy, where the nested AUC fell to 0.561, an expected consequence of chemotherapy altering both tumour metabolism and imaging appearance before surgery.

Why does a modest result merit attention? Because the study is a methodological exemplar in a field awash with overclaimed performance. Many published radiomics models report dazzling AUCs that evaporate under external validation, largely because of leakage, small samples and optimistic model selection. By embedding every preprocessing and selection decision inside nested cross-validation, quantifying selection optimism explicitly, reporting calibration alongside discrimination and publishing decision-curve analyses, the Chinese team has provided a template for how multimodal prediction models should be evaluated. The result, a moderate AUC of roughly 0.71 to 0.80 with imperfect calibration, is likely a far more accurate reflection of what such models can currently achieve than the near-perfect figures that populate the literature.

The authors are explicit that this work should be regarded as a proof-of-concept framework rather than a clinical tool. The cohort was small, retrospective and drawn from a single centre, the First Affiliated Hospital of Henan University of Science and Technology in Luoyang, and the study was approved by that institution’s medical ethics committee with the requirement for written informed consent waived owing to the retrospective design. Before any clinical application, prospective multicentre external validation is required, including direct comparison with routine visual PET/CT assessment of the axilla. If such validation confirms even the moderate performance observed here, the payoff could be substantial: a non-invasive preoperative estimate of nodal status could help select patients for sentinel biopsy versus more extensive axillary surgery, spare node-negative women unnecessary procedures and inform neoadjuvant treatment decisions. For now, the study’s greatest contribution may be cultural, demonstrating that in the race to bring artificial intelligence into cancer staging, honesty about uncertainty is not a weakness but the very foundation of trustworthy medicine.

Subject of Research: Preoperative prediction of axillary lymph-node metastasis in invasive breast cancer using a leakage-controlled multimodal machine-learning model based on 18F-FDG PET/CT radiomics, deep-learning features, metabolic parameters and haematological inflammation-nutrition indices

Article Title: A leakage-controlled multimodal model integrating 18F-FDG PET/CT radiomics, deep-learning features, metabolic parameters and haematological inflammation–nutrition indices for preoperative prediction of axillary lymph-node metastasis in invasive breast cancer

Article References: A leakage-controlled multimodal model integrating 18F-FDG PET/CT radiomics, deep-learning features, metabolic parameters and haematological inflammation–nutrition indices for preoperative prediction of axillary lymph-node metastasis in invasive breast cancer. (n.d.). https://doi.org/10.1186/s12880-026-02948-2

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02948-2

Keywords: breast cancer, axillary lymph-node metastasis, 18F-FDG PET/CT, radiomics, deep learning, machine learning, nested cross-validation, data leakage, biomarkers, cancer staging, predictive medicine, neoadjuvant therapy