Academic radiology · 2026
Vision Transformer for Pretreatment Prediction of Pathologic Complete Response in Multi-center Breast DCE-MRI.
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.
Fridman Naomi, Solway Bubby, Fridman Tomer, Barnea Itamar, Goldstein Anat
Аннотация
RATIONALE AND OBJECTIVES: Pretreatment prediction of pathologic complete response (pCR) is clinically important for guiding neoadjuvant chemotherapy (NAC), but robust multi-center benchmarks with standardized evaluation protocols remain limited. We evaluated a Vision Transformer (ViT) approach for pCR prediction from breast dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI).
MATERIALS AND METHODS: We used BreastDCEDL, a public multi-center dataset comprising 2070 patients from the I-SPY1, I-SPY2, and Duke cohorts with predefined splits. Pre-contrast, early post-contrast, and late post-contrast phases were mapped to RGB channels for transfer learning from ImageNet-pretrained backbones. The held-out test set (n=175) included 32 pCR+ in I-SPY2 (99), 8 in Duke (41), and 12 in I-SPY1 (35). For I-SPY2, ViT scores were combined with clinical variables (age, HR/HER2 status, tumor volume, genomic score, treatment arm) via a Random Forest classifier.
RESULTS: The ViT achieved AUC 0.72 (95% CI 0.64-0.80), sensitivity 0.27 (95% CI 0.15-0.40), and specificity 0.96 (95% CI 0.92-0.99). Performance varied across cohorts (AUC: I-SPY2 0.78, 95% CI 0.68-0.86, I-SPY1 0.68, 95% CI 0.46-0.89, Duke 0.54, 95% CI 0.27-0.80). In I-SPY2, clinical variables increased AUC to 0.85 overall (n=99), a difference that was not statistically significant (DeLong P=0.06).
CONCLUSION: Cross-cohort variation highlights challenges in developing generalizable pretreatment response models under real-world heterogeneity. The high specificity supports potential utility for identifying non-responders, enabling earlier transition to alternative strategies. Code, data, and predefined splits are publicly available, establishing a reproducible benchmark.
MATERIALS AND METHODS: We used BreastDCEDL, a public multi-center dataset comprising 2070 patients from the I-SPY1, I-SPY2, and Duke cohorts with predefined splits. Pre-contrast, early post-contrast, and late post-contrast phases were mapped to RGB channels for transfer learning from ImageNet-pretrained backbones. The held-out test set (n=175) included 32 pCR+ in I-SPY2 (99), 8 in Duke (41), and 12 in I-SPY1 (35). For I-SPY2, ViT scores were combined with clinical variables (age, HR/HER2 status, tumor volume, genomic score, treatment arm) via a Random Forest classifier.
RESULTS: The ViT achieved AUC 0.72 (95% CI 0.64-0.80), sensitivity 0.27 (95% CI 0.15-0.40), and specificity 0.96 (95% CI 0.92-0.99). Performance varied across cohorts (AUC: I-SPY2 0.78, 95% CI 0.68-0.86, I-SPY1 0.68, 95% CI 0.46-0.89, Duke 0.54, 95% CI 0.27-0.80). In I-SPY2, clinical variables increased AUC to 0.85 overall (n=99), a difference that was not statistically significant (DeLong P=0.06).
CONCLUSION: Cross-cohort variation highlights challenges in developing generalizable pretreatment response models under real-world heterogeneity. The high specificity supports potential utility for identifying non-responders, enabling earlier transition to alternative strategies. Code, data, and predefined splits are publicly available, establishing a reproducible benchmark.