BMC medical imaging · 2026
AI-based multimodal fusion for preoperative prediction of breast microcalcifications: combining mammography and tomosynthesis.
СтатьяНазвание и аннотация: Перевод в очереди
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.
Wei Junting, Li Jie, Pan Ruixin, Cao Mingna, Ma Chengcheng, Xie Zongyu, Ma Yichuan, Gao Zhizhen
Аннотация
BACKGROUND: This study aimed to develop artificial intelligence (AI) models based on mammography (MG) and digital breast tomosynthesis (DBT) for non-invasive preoperative differentiation of benign and malignant breast microcalcifications.
METHODS: We retrospectively analyzed 190 patients with pathologically confirmed breast microcalcifications, who were divided into training and test sets at a ratio of 7:3. Radiomics features and deep learning scores extracted using ResNet-50 were derived from MG and DBT images. After feature selection procedures, single-modality models and an integrated model (MD model) were constructed.
RESULTS: The area under the curves (AUCs) of the BI-RADS model, DBT model, MG model, and MD model in the training set were 0.805, 0.879, 0.896, and 0.939. In the test set, the corresponding AUCs were 0.809, 0.768, 0.793, and 0.835, respectively. The MD model demonstrated significantly better diagnostic performance compared with the other models (DeLong test, P < 0.05). SHapley Additive exPlanations (SHAP) analysis revealed that the deep learning scores (DL scores) from the MG craniocaudal (CC) view and radiomics features reflecting textural complexity and spatial heterogeneity were the most influential features in model prediction.
CONCLUSIONS: The MD model integrating radiomics and deep learning scores from MG and DBT improved the diagnostic accuracy for differentiating benign and malignant breast microcalcifications, outperforming traditional methods. It provided a valuable non-invasive preoperative tool with the potential to optimize clinical decision-making and reduce unnecessary invasive procedures.
METHODS: We retrospectively analyzed 190 patients with pathologically confirmed breast microcalcifications, who were divided into training and test sets at a ratio of 7:3. Radiomics features and deep learning scores extracted using ResNet-50 were derived from MG and DBT images. After feature selection procedures, single-modality models and an integrated model (MD model) were constructed.
RESULTS: The area under the curves (AUCs) of the BI-RADS model, DBT model, MG model, and MD model in the training set were 0.805, 0.879, 0.896, and 0.939. In the test set, the corresponding AUCs were 0.809, 0.768, 0.793, and 0.835, respectively. The MD model demonstrated significantly better diagnostic performance compared with the other models (DeLong test, P < 0.05). SHapley Additive exPlanations (SHAP) analysis revealed that the deep learning scores (DL scores) from the MG craniocaudal (CC) view and radiomics features reflecting textural complexity and spatial heterogeneity were the most influential features in model prediction.
CONCLUSIONS: The MD model integrating radiomics and deep learning scores from MG and DBT improved the diagnostic accuracy for differentiating benign and malignant breast microcalcifications, outperforming traditional methods. It provided a valuable non-invasive preoperative tool with the potential to optimize clinical decision-making and reduce unnecessary invasive procedures.