Evaluating axillary lymph node metastasis risks in breast cancer patients via Semi-ALNP: a multicenter study.

Background: Accurately evaluating axillary lymph nodes (ALNs) is essential for guiding both staging and treatment strategies in breast cancer (BC) patients. Currently, traditional pathological staging methods still rely on invasive biopsies or surgeries. This study aimed to construct, evaluate, and validate a semisupervised classifier utilizing radiomic and machine learning (ML) techniques to noninvasively identify axillary nodal disease.

Methods: Data from 4191 ALNs in 494 patients with invasive BC were retrospectively analyzed at the Second Xiangya Hospital of Central South University between January 31, 2016, and July 31, 2024, including a labeled cohort (214 patients, 1769 ALNs, divided into ultra-low and ultra-high risk groups) and an unlabeled cohort (280 patients, 2422 ALNs). Regions of interest (ROIs) were segmented, and CT radiomic features were extracted. 11 supervised learning models were built on the basis of labeled ALNs, and pseudolabels (low-risk, high-risk groups) were assigned to unlabeled ALNs. Seven ML algorithms developed semisupervised multiclassifiers on the basis of the predicted probabilities for 4191 ALNs. For multicenter validation, additional data were collected from the First People's Hospital of Chenzhou City, the First People's Hospital of Changde City, and the First People's Hospital of Xiangtan City. The best-performing multiclassifier was evaluated in two independent multicenter cohorts: 212 clinically node-positive (cN+) patients who underwent core needle biopsy (CNB) or fine needle aspiration (FNA), and 450 clinically node-negative (cN0) patients. The research was registered at www.isrctn.com with registration number ISRCTN54288903.

Findings: The supervised multilayer perceptron (MLP) model, built from labeled ALNs, exhibited excellent classification performance, with an area under the curve (AUC) of 0.959 (95% CI: 0.937-0.981), a sensitivity of 0.899, and a specificity of 0.932. Pseudolabels for the unlabeled ALNs were generated via this model, and the semisupervised MLP multiclassifier (Semi-ALNP) was constructed by combining the labeled and unlabeled data. The AUCs for predicting nodal metastases were 0.906 (95% CI: 0.894-0.917), 0.936 (95% CI: 0.928-0.945), 0.948 (95% CI: 0.940-0.956), and 0.955 (95% CI: 0.946-0.965) for the ultra-low risk, low-risk, high-risk, and ultra-high risk groups, respectively. Validation in both the biopsy and cN0 cohorts revealed strong diagnostic performance: in the biopsy cohort, the model achieved a false negative rate (FNR) of 1.21%, a false positive rate (FPR) of 14.89%, a sensitivity of 98.79%, and a specificity of 85.11%; in the cN0 cohort, the FNR was 8.33%, the FPR was 9.94%, the sensitivity was 91.67%, and the specificity was 90.06%.

Interpretation: Semi-ALNP, which is based on the MLP algorithm, has high accuracy in assessing the statuses of ALNs across all types of BC patients. It is particularly effective for identifying high-risk patients with ALN metastasis, which can help guide personalized treatment decisions. Future prospective studies are planned to further validate the clinical utility of this approach in real-world settings.

Funding: This study was funded by the Science and Technology Innovation Program of Hunan Province (Grant No. 2021SK2026) and the Innovation Platform and Talent Plan of Hunan Province (2023SK4019). Funding sources were not involved in the study design, data collection, analysis and interpretation, writing of the report, or decision to submit the article for publication.