Frontiers in artificial intelligence · 2026
BCMM: gated bidirectional cross-attention fusion of breast MRI and clinical features for predicting RCB response in the I-SPY1 cohort.
СтатьяНазвание и аннотация: Перевод пока не подготовлен
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Yang Mu, Hu Junfeng
Аннотация
OBJECTIVE: To develop a multimodal predictive framework that integrates breast magnetic resonance imaging (MRI) and structured clinical features for predicting residual cancer burden (RCB) following neoadjuvant therapy.
METHODS: The proposed Breast Cancer Multi-source Multi-scale Model (BCMM) was evaluated on the I-SPY1 cohort (n = 201), incorporating eight structured clinical variables and 2.5D MRI inputs. The model consists of a clinical feature encoding branch (MLP with squeeze-and-excitation), a Vision Transformer (ViT)-based imaging encoder with multi-scale feature aggregation, and a gated bidirectional cross-attention fusion module. Model performance was assessed using stratified five-fold group cross-validation at the patient level. Additionally, bootstrap resampling was applied to pooled out-of-fold (OOF) predictions to estimate 95% confidence intervals.
RESULTS: Across the five folds, BCMM achieved a mean AUC of 0.826 ± 0.076 and a mean AUPRC of 0.926 ± 0.028. On pooled out-of-fold predictions, BCMM achieved an AUC of 0.795, an AUPRC of 0.895, an accuracy of 0.677, a balanced accuracy of 0.727, a macro-F1 score of 0.662, and an MCC of 0.406.
CONCLUSION: BCMM showed promising internal cross-validated discrimination in this single-cohort study. Independent external validation is required before conclusions regarding robustness or clinical utility can be made.
METHODS: The proposed Breast Cancer Multi-source Multi-scale Model (BCMM) was evaluated on the I-SPY1 cohort (n = 201), incorporating eight structured clinical variables and 2.5D MRI inputs. The model consists of a clinical feature encoding branch (MLP with squeeze-and-excitation), a Vision Transformer (ViT)-based imaging encoder with multi-scale feature aggregation, and a gated bidirectional cross-attention fusion module. Model performance was assessed using stratified five-fold group cross-validation at the patient level. Additionally, bootstrap resampling was applied to pooled out-of-fold (OOF) predictions to estimate 95% confidence intervals.
RESULTS: Across the five folds, BCMM achieved a mean AUC of 0.826 ± 0.076 and a mean AUPRC of 0.926 ± 0.028. On pooled out-of-fold predictions, BCMM achieved an AUC of 0.795, an AUPRC of 0.895, an accuracy of 0.677, a balanced accuracy of 0.727, a macro-F1 score of 0.662, and an MCC of 0.406.
CONCLUSION: BCMM showed promising internal cross-validated discrimination in this single-cohort study. Independent external validation is required before conclusions regarding robustness or clinical utility can be made.