TABLE 4.

Clinical studies on the application of AI in breast cancer screening, feature characterization, and surveillance.

Study	Imaging modality	Application	Key findings
Kim HE et al.[ 30d ]	MG	Screening	The AI algorithm showed better diagnostic performance in breast cancer detection compared with radiologists (AUC: 0.940 vs 0.810). Radiologists’ performance improved (0.881) when aided by AI.
Rodriguez‐Ruiz A et al.[ 41 ]	MG	Screening	AI can be used to automatically preselect screening items to reduce the reading workload for breast cancer screening.
Jiang Y et al.[ 42 ]	US	Screening	The deep learning‐based CAD system helped radiologists improve screening accuracy (AUC: 0.848 vs. 0.828).
Qi X et al.[ 43 ]	US	Screening	Mt‐Net and Sn‐Net were proposed to identify malignant tumours and recognize solid nodules in a cascade manner, with an AUC of 0.982 for Mt‐Net and 0.928 for Sn‐Net.
Illan IA et al.[ 44 ]	MRI	Screening	Independent component analysis was used to extract data‐driven dynamic lesion characterizations to address the challenges of nonmass‐enhancing lesion detection and segmentation.
Chougrad H et al.[ 46a ]	MG	Classification of malignant and benign lesions	A CAD system based on a dCNN model to classify benign and malignant lesions (AUC = 0.99, accuracy: 98.23%).
Akselrod‐Ballin A et al.[ 45a ]	MG	Classification of malignant and benign lesions	An algorithm combining ML and DL approaches was used to classify benign and malignant masses (AUC = 0.91), with a level comparable to radiologists.
Ciritsis A et al.[ 46b ]	US	Classification of lesions according to BI‐RADS catalog	dCNNs may be used to mimic human decision‐making in classifying BI‐RADS 2–3 versus 4–5 (accuracy: 93.1% vs 91.6 ± 5.4%).
Han S et al.[ 46c ]	US	Classification of malignant and benign lesions	The proposed deep learning‐based method had great discriminating performance in a large dataset for classifying benign and malignant lesions (Accuracy: 90%, sensitivity: 0.86, specificity: 0.96).
Fujioka T et al.[ 30c ]	US	Classification of malignant and benign lesions	Deep learning with CNN showed a high diagnostic performance to discriminate between benign and malignant breast masses on ultrasound (AUC  =  0.913 vs 0.728−0.845).
Jiang Y et al.[ 62 ]	MRI	Classification of malignant and benign lesions	The AI system improved radiologists’ performance in differentiating benign and malignant breast lesions on MRI (ΔAUC = +0.05).
Zhou BY et al.[ 47 ]	US	Prediction of molecular subtypes	The multimodal US‐based ACNN model performed better than the monomodal or dual‐modal model in predicting four‐classification, five‐classification molecular subtypes and identifying TNBC from non‐TNBC.
Ma M et al.[ 48 ]	US, MG	Prediction of molecular subtypes	The interpretable machine learning model could help clinicians and radiologists differentiate between breast cancer molecular subtypes, and the decision tree model performed best in distinguishing TNBC from other breast cancer subtypes (AUC = 0.971).
Ha R et al.[ 63 ]	MRI	Prediction of molecular subtypes	A CNN algorithm was developed to predict the molecular subtype of breast cancer based on features on breast MRI images (accuracy: 70%, AUC = 0.853).
Zhou LQ et al.[ 49a ]	US	Prediction of lymph node metastasis	The best‐performing CNN model, Inception V3, predicted clinically negative axillary lymph node metastasis with higher sensitivity (85% vs 73%) and specificity (73% vs 63%) than radiologists.
Zhang Q et al.[ 49b ]	US	Prediction of lymph node metastasis	A computer‐assisted method using dual‐modal features extracted from real‐time elastography and B‐mode ultrasound was valuable for discrimination between benign and metastatic lymph nodes (AUC = 0.895).
Zheng X et al.[ 49c ]	US	Prediction of lymph node metastasis	DL radiomics of conventional ultrasound and shear wave elastography combined with clinical parameters yielded the best diagnostic performance in predicting disease‐free axilla and any axillary metastasis in early‐stage breast cancer (AUC = 0.902).
Ha R et al.[ 50a ]	MRI	Prediction of lymph node metastasis	The feasibility of using CNN models to predict lymph node metastasis based on MRI images was proved (accuracy: 84.3%).
Yu Y et al.[ 50b ]	MRI	Prediction of lymph node metastasis	A multiomic signature incorporating key radiomic features extracted by machine learning random forest algorithm, clinical and pathologic characteristics, and molecular subtypes could preoperatively identify patients with axillary lymph node metastasis in early‐stage invasive breast cancer (AUC = 0.90 and 0.91 in the training and external validation sets, respectively).
Jiang L et al.[ 51b ]	MRI	Tumour microenvironment revelation; prognosis prediction	Peritumoral heterogeneity correlated with metabolic and immune abnormalities in TNBC; radiomic features reflecting peritumoral heterogeneity indicated TNBC prognosis.
Arefan D et al.[ 51a ]	MRI	Tumour microenvironment revelation	Breast MRI‐derived radiomics extracted by a radio‐genomics approach and machine learning models were associated with the tumour's microenvironment in terms of the abundance of several cell types.
Kavya R et al.[ 54a ]	MRI	Prediction of NAC response	A CNN model combining pre‐ and postcontrast DCE‐MRI images was proposed to predict which NAC recipients will achieve pathological complete response (AUC = 0.77).
Gu J et al.[ 54b ]	US	Prediction of NAC response	Deep learning radiomics models were proposed to stepwise predict the response to NAC at different NAC time points.