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

Abstract

The status of the axillary lymph nodes (ALN) guides staging and the extent of axillary surgery in invasive breast cancer, yet confirmation still depends on sentinel lymph-node biopsy or axillary dissection. This exploratory study developed and internally evaluated, by cross-validation, a model combining five complementary streams of information available at staging ( 18 F-FDG PET/CT radiomics, deep-learning (DL) imaging features, volumetric metabolic parameters, routine clinicopathological variables and blood-count-derived inflammation–nutrition indices) within a machine-learning pipeline designed to avoid information leakage. Sixty-three women with histologically confirmed invasive breast carcinoma who underwent baseline 18 F-FDG PET/CT at a single centre were analysed retrospectively (42 node-positive, 21 node-negative); 40 (63%) subsequently received neoadjuvant systemic therapy. Features were extracted from the tumour volume delineated on co-registered PET and CT: 2,167 IBSI-compliant radiomics features, 12,288 ResNet-50 transfer-learning features, 11 volumetric metabolic parameters, 7 clinicopathological variables and 7 inflammation–nutrition indices. Median imputation, variance filtering, standardisation, univariate AUC screening, collinearity pruning and LASSO selection were fitted only within the training folds of repeated stratified five-fold cross-validation (10 repeats). Five single-block models and three fusion strategies (stacking, feature concatenation and a parameter-free probability mean) were compared across four base learners, and the complete choice of block, fusion strategy, learner and decision threshold was repeated inside a nested cross-validation. Uncertainty was estimated by patient-level bootstrap, models were compared by paired bootstrap, and calibration, decision curves, the incremental value of imaging over clinicopathological variables, the handling of missing data and robustness to contour perturbation were assessed. The nested cross-validated AUC of the whole modelling procedure was 0.711 (95% CI 0.609–0.805). The probability-mean fusion, the most frequently selected strategy, reached 0.795 (0.678–0.891) on the same splits, indicating a model-selection optimism of about 0.08; its sensitivity and specificity at training-fold thresholds were 0.80 and 0.60, and calibration was imperfect (slope 1.53). Single-block AUCs ranged from 0.590 (inflammation–nutrition; 95% CI including 0.5) to 0.750 (DL). Adding PET/CT-derived information to the clinicopathological variables (AUC 0.681, 0.544–0.805) increased the AUC by 0.107 (− 0.023 to 0.247; P  = 0.12), and net benefit exceeded that of a treat-all strategy only at threshold probabilities of about 0.4 or higher. Discrimination was poor in patients who received neoadjuvant therapy (nested AUC 0.561). Under a stringent one-voxel contour perturbation, 23% of radiomics features had an intraclass correlation coefficient of at least 0.75, and restricting the radiomics block to these features did not reduce performance. In this small single-centre retrospective cohort, combining PET/CT-derived and routinely available clinical information under leakage-controlled, nested cross-validation gave moderate discrimination of axillary nodal status, with imperfect calibration and no statistically demonstrable incremental value of imaging over clinicopathological variables. The work should be regarded as a proof-of-concept framework; prospective multicentre external validation, including comparison with routine PET/CT axillary assessment, is required before any clinical application.