TALON: a multi-task network for axillary lymph-node segmentation and source-cohort classification on breast DCE-MRI

Purpose This study developed TALON (Target-Aware Lesion-Oriented Network) for pre-selected axillary target-node segmentation and exploratory invasive breast cancer vs. idiopathic granulomatous mastitis source-cohort classification on single-phase dynamic contrast-enhanced MRI (DCE-MRI). Methods This retrospective study included 103 patients, 104 side-specific cases, and 1,367 axial slices. Expert target-node annotations supported segmentation evaluation. TALON and a Mask-guided Multi-task U-Net were compared using five patient-grouped holdouts with 60%/20%/20% training, validation, and test partitions. Model selection, segmentation thresholds, and Platt calibration were performed using validation data. Exploratory analyses examined component removal, age confounding, and the association and predictive value of adding TALON probabilities to age. Results Mean case-level AUROC was 0.905 ± 0.047 for TALON and 0.838 ± 0.113 for the comparator; mean slice-level Dice was 0.530 ± 0.057 and 0.424 ± 0.085, respectively. Paired AUROC and calibrated Brier-score difference intervals included zero in every repeat. The training-fitted age-only mean AUROC was 0.945 ± 0.046. In post-hoc multivariable analysis, TALON probability was associated with cohort membership after adjustment for age (OR, 7.615 per SD; 95% CI, 2.795–20.748). A separate validation-fitted age-plus-TALON model achieved mean test AUROC 0.965 ± 0.022. Relative to its matched age-only model, the AUROC difference was 0.020 (95% CI, −0.015 to 0.063) and the Brier difference was −0.031 (95% CI, −0.052 to −0.009). TALON's pointing-game fraction was zero across seeds, and all paired single-seed ablation intervals included zero. Conclusion TALON supports further study of target-node segmentation and the use of complementary imaging information alongside age. The findings remain an internal proof of concept and do not establish age-independent disease discrimination, target-node-driven classification, stable superiority, nodal metastatic status, or clinical utility. Component and training-schedule benefits require multi-seed evaluation.